AI Meets the
Physical World
Episode 001
Models, megawatts & useful work
Ben Lyddane
BTUsandGPUs.com
AI Is a Five-Layer Cake
Jensen Huang’s framework · read from the foundation up
Three Ways AI Can Help
Enable new work—or improve a process you already have
Public Concern About AI
Original survey visuals · different questions and populations
U.S. / China comparison · October 2025
Net favorability · original deck graphic
2026 Pew update: concern exceeds excitement in all four age groups.
Context Windows Have Grown
Original 2018–2025 chart · context size is not the same as useful recall
The Current Model Landscape
Representative releases · checked September 8, 2026
OpenAI
GPT‑6 Astra
Reasoning · coding · computer use
Anthropic
Claude Fable 5.1
Coding · vision · long-running work
Google
Gemini 3.8 Flash
Fast multimodal tasks
xAI
Grok 4.20
Reasoning · tool use
Task qualityContext + toolsLatencyTotal cost
Four Lenses on the Same System
Architecture ≠ modality ≠ function ≠ deployment
Capability and Compute
Measure capability, then follow the hardware it needs
Benchmarks: Tools Change the Result
Original HLE comparison · tools and test conditions change the result
Accuracy and Test-Time Cost
ARC-AGI-1 · read both axes
A Harder Test Reveals Different Gaps
ARC-AGI-2 · benchmark versions are not interchangeable
Compare the full setup
Model + reasoning budget
Harness + tools + test split
The series now includes ARC-AGI-3
Interactive tasks introduce a different evaluation setting.
The Physical Limits Behind AI
Making the chip—and moving its heat
MAKE THE CHIPTin plasma → EUV mirrors → wafer
REJECT THE HEATCompute → heat transport → radiator
Fast Is Not the Same as Correct
A model-specific specialized-inference demonstration
17,000tokens / second / user
HC1Llama 3.1 8B2.5 kW server
One Megawatt Is a Mechanical Problem
An illustrative steady-state IT load
The $7 Trillion Buildout
Investment and demand are related—but different measures
Where Projects Cluster
The original construction map provides a geographic snapshot
A pipeline is not delivered capacity
POWER · LAND · FIBER
DELIVERY · COMMISSIONING
Original graphic: March 2026. Counts are not refreshed to today.
Every Requirement Crosses Several Handoffs
Technical choices, delivery constraints and operations must agree
What Buyers Need From Suppliers
Reliability, specification compliance and delivery
Perform.
Fit.
Arrive.
A correct component must work as part of the complete system.
Follow the Heat—And Both Water Loops
Capturing heat at the chip is different from rejecting it outside
Specify the Temperature Interface
Facility-water classes are not universal GPU inlet temperatures
Maximum facility supply
113°F
Named IT inlet example
NVIDIA Vera Rubin NVL72
The location of the temperature matters.
When Does the Compressor Turn On?
Example: dry cooler + chiller plant, with no evaporative assist
| Outdoor air | 80.6°F inlet | 95°F inlet | 113°F inlet |
|---|
| 68°F | CHILLER ON | DRY MODE | DRY MODE |
| 86°F | CHILLER ON | CHILLER ON | DRY MODE |
| 95°F | CHILLER ON | CHILLER ON | DRY MODE |
| 104°F | CHILLER ON | CHILLER ON | CHILLER ON |
What the Compressor Actually Does
The compressor raises refrigerant pressure and temperature so heat can move outdoors
Agents and Useful Work
Back to the application layer: turn answers into completed work
The Agent Is More Than the Model
A tool call belongs inside a controlled application loop
Interest in Personal Agents
OpenClaw and the shift beyond a chat interface
Interest is not reliability
GitHub stars measure attention.
They do not count dependable deployments.
The reservation story
Seven people?
Seven o’clock?
The constraint matters.
Model Context Protocol
A shared connection between AI applications, tools and data
MCP standardizes the connection
Servers can still use APIs underneath.
The application controls actions
Permissions and enforcement remain part of the system.
MCP Connects the Assistant to Real Systems
The underlying tools and APIs still do the work
From Tools to a Working Process
A brief example from my work
Project documents
→Equipment data
→Team workflow
Map the Work Before Automating It
A detailed equipment-sales workflow is one industry example
Choose one repeated handoff with a clear input, result and reviewer.
A Bid Record With Evidence Behind It
Intake → extraction → source check → reviewed record
Example: Extract, Match, Verify
Three links must survive the handoff
Source specification
→Extracted field
→Catalog candidate
| Tag | Extracted item | Qty | Catalog match | Review |
|---|
| AHU-1 | Air handling unit | 1 | Candidate A | Airflow |
| CH-1 / 2 | Water-cooled chiller | 2 | Candidate B | Duty |
| EF-1 / 2 / 3 | Exhaust fan | 3 | Candidate C | Pressure |
A Drawing Contains More Than Text
Symbols, schedules, annotations and revisions must agree
Manufacturer × Category
Original analytics screenshot · the selected project sample
Your First Useful Workflow
Build a small test you can actually judge
