Hello, I’m Joel.
My experience includes OpenAI and, most recently, Prometheus. I work across the AI/ML stack, from model architecture, training recipes, parallelism, and inference behavior to accelerators and networks, then power, data centers, capacity products, allocation policy, and the finance plans behind them.
I came to this work through forecasting, supply chains, data center strategy, TPU and GPU planning, and inference scaling. In practice, I turn AI workload demand into the systems that run it. That means understanding what a model needs, helping shape what teams build, and asking whether installed supply is becoming useful capacity.
Begin with the workload, not the purchase order.
The evidence behind the thesis
Each role moved closer to the model while keeping the same planning discipline: understand demand, shape the system, build the supply, and operate the resulting capacity.
Prometheus · 2026 - now
Helping establish compute planning early in Prometheus's infrastructure build, connecting emerging model workloads to accelerator, network, power, data center, and capacity-product decisions.
OpenAI · 2024 - 2026
Founded the compute-planning function and later led inference planning. Built the planning discipline as inference demand expanded from roughly 250 million to 1 billion weekly active users and the accelerator estate grew to millions of compute chips.
Google · 2022 - 2024
Led data center planning for large-scale TPU deployments. Authored Resource Aware Power Planning and the solver used to orchestrate power allocation across AI product areas.
Worked across compute and infrastructure planning.
During the COVID-19 demand shock, worked within the response-planning team to translate sudden e-commerce growth into capacity and supply-chain decisions, helping meet customer needs while keeping the network intact.
Built the analytical and experimental foundations that still shape how I approach uncertain systems.
Selected public record
A compact record of the work that is easiest to inspect: peer-reviewed research, collective model releases, and conversations about the compute layer.
Applied Mechanics and Materials · DOI 10.4028/www.scientific.net/AMM.592-594.630
International Conference on Innovations in Engineering and Technology · DOI 10.15242/IIE.E1213517
An ICLR workshop paper evaluates GPT-5 mini for spatial reasoning. I am listed among the contributors to the public GPT-5 release, making this a downstream example of collective model work entering research practice.
Infrastructure work is collective by nature. These OpenAI release pages publicly acknowledge my contribution alongside the teams that built them.
SDxCentral
Yotta 2025
Xcelerated Compute 2026
2026 program
A small protected lab
I am building Compute Academy as a visual, hands-on field guide to the complete model system: learning, Transformers, pretraining, post-training, accelerators, distributed execution, inference, and fleet operations. It starts gently, then keeps going until you can train something, serve it, and reason about the machine underneath it.
If hearing GPUs go brr is your thing, hit me up for access.