Acquire
Understand CPUs, GPUs, memory, storage, power, thermals, compatibility, and cost.
Founding cohort · Silicon Valley Live garage lab
Build AI compute.
Not just AI apps.
Source, assemble, boot, benchmark, and optimize a real multi-vendor AI system in a working garage lab. Learn the physical infrastructure behind every model you use.
Hands-onEvery session
Multi-vendorAMD · NVIDIA · Intel
Small cohortBuilt for builders
The premise
Every AI product runs on physical machines. Almost nobody in tech has actually built one.
Train Your Dragon closes that gap. This is not another software tutorial. It is a guided build experience spanning silicon, servers, power, cooling, networking, Linux, runtimes, inference, observability, and economics.
The experience
Learn infrastructure by touching it.
You will move through the same chain of decisions that real infrastructure teams face, from component selection to workload performance.
Assemble
Build a server from components and learn what each connector, rail, fan, and slot actually does.
Boot
Configure firmware, install Linux, provision drivers, and bring the hardware to life.
Deploy
Launch models with production inference software and expose a working endpoint.
Benchmark
Measure latency, throughput, utilization, memory pressure, power, and system behavior.
Optimize
Find bottlenecks, change the system, and prove whether performance actually improved.
Curriculum
The complete AI compute stack.
Each module connects the hardware beneath the rack to the software serving the model.
System architecture
Platform topology, PCIe, memory hierarchy, NUMA, storage, power delivery, airflow, and thermal limits.
Build and bring-up
Assembly, firmware, BIOS, Linux provisioning, drivers, device discovery, and validation.
Networks and storage
Ethernet, topology, bandwidth, latency, local and shared storage, and data movement.
AI software stack
Containers, CUDA and ROCm concepts, PyTorch, model formats, serving engines, and APIs.
Inference performance
Prefill, decode, batching, KV cache, quantization, TTFT, throughput, and tail latency.
Observability and economics
Telemetry, utilization, power, failure diagnosis, capacity, cost per token, and optimization tradeoffs.
The garage is the classroom
“The fastest way to understand AI infrastructure is to build it, break it, and make it faster.”
Small cohort. Real equipment. No simulated lab.
What you leave with
A working system and a new mental model.
You will understand how hardware, software, workload behavior, and economics interact, not as abstract concepts but as a system you personally assembled and operated.
Apply to build yours →01
Hardware fluency
Speak confidently about components, constraints, compatibility, and tradeoffs.
02
Operational confidence
Bring up a machine, diagnose failures, and navigate the Linux and GPU stack.
03
Performance intuition
Connect model behavior to compute, memory, networking, power, and cost.
04
A demonstrable build
Leave with benchmarks, architecture notes, and a system story you can show.
Who this is for
Curious builders, not passive students.
You do not need to be a hardware expert. You do need curiosity, persistence, and a willingness to get your hands dirty.
Move below the API layer
Understand what your models and applications demand from the machines beneath them.
Make better infrastructure decisions
Evaluate compute choices, architecture tradeoffs, vendors, and operating costs.
Build practical credibility
Develop a rare, tangible foundation across systems, hardware, and AI infrastructure.
COHORT 01 APPLICATION
Build something real.
Required fields are marked with an asterisk.
FAQ
Before you enter the lab.
More cohort details will be shared with selected applicants.
Do I need prior hardware experience?+
No. The program is designed to make hardware and systems concepts approachable. Basic technical comfort and strong curiosity are more important.
Is this a coding bootcamp?+
No. Code is part of the experience, but the focus is the full AI compute system, including hardware, Linux, networking, runtimes, performance, power, and cost.
Will I work with real equipment?+
Yes. The defining feature of Train Your Dragon is hands-on work with actual servers, components, accelerators, cables, tools, and software.
Where will the program take place?+
The founding cohort is planned for the San Jose and greater Bay Area. Exact dates and venue details will be shared with shortlisted applicants.
Does applying guarantee a place?+
No. Cohort size is intentionally limited to preserve the hands-on format. Applications will be reviewed for fit, commitment, and cohort balance.