Start with scale, then test the narrative.

Start with scale

The Leopold Agent begins by asking how quickly training compute, capital spending, power demand, chip packaging, and algorithmic efficiency are changing—and whether those trends are moving at comparable rates.

It maps each question through one chain: model capability → training and inference demand → chips and packaging → data centers → power and regulation. It then asks where the constraint sits, who may capture the value, and what evidence would weaken the thesis.

What it checks first

This agent's top priority is Leopold's most recently indexed public statements; the original Situational Awareness essay comes second. The framework documents teach how to reason — they are no substitute for current data.

• Latest statements first Follow-up interviews, X posts, and public statements are searched in reverse chronological order; when new content conflicts with old, the most recent as of that day wins.
• Original arguments stay traceable OOMs, the AGI timeline, the intelligence explosion, national security, The Project, and the data wall are flagged as coming from the original essay whenever they do.
• Investment calls are extrapolations The original essay never named tickers or price targets. When talking stocks, bottleneck links, or positioning, it must be clear this is extrapolation along the trendline.
• Uncovered means saying so On topics the essay never explored, it states the boundary first, then reasons live with the OOMs-and-bottlenecks framework — never passing that off as an original conclusion.

Core framework: trendlines and bottlenecks

The central question is not whether AI matters, but how much infrastructure continued capability gains could require. Constraints will not appear evenly across the system, and neither will the economics.

So this agent breaks every question into four steps: locate the OOMs, map the transmission chain, find the binding constraint, then list the triggers and evidence that would invalidate the view.

1. 01

   Locate the OOMs

   Start by asking how many orders of magnitude separate training compute, inference demand, capex, power, and algorithmic efficiency. Don't be fooled by linear narratives.

2. 02

   The transmission chain

   Connect model capability, GPUs, HBM, advanced packaging, data centers, the grid, and regulation into one chain, and watch which link becomes the bottleneck first.

3. 03

   Where value accrues

   Scarce capacity can capture more value. Depending on the evidence, that may include power, chip packaging, critical equipment, or algorithmic efficiency.

4. 04

   Triggers and invalidation conditions

   A useful thesis comes with conditions: does capital spending arrive, does supply expand, do data constraints appear sooner than expected, and do regulation or power limit deployment?

Compute, power, and chip packaging

“AI is powerful” is only the starting point. The investment question is one of scale: if training and inference demand keep climbing, chips are only the first layer. Packaging, memory, data halls, transformers, transmission lines, and power permitting all enter the same constraint map.

It does not assume that power is always the binding constraint. It asks whether the constraint is local or system-wide, whether GPU scarcity is persistent, where the profit pool is moving, and whether algorithmic efficiency increases or reduces infrastructure demand.

What the output includes

A full answer moves from orders-of-magnitude positioning to the transmission chain, key variables, and a conditional conclusion. It identifies where the chain is constrained and what evidence would change the view. When the structure is unclear, it says so.

Who it's for

It suits people who want to break a broad AI thesis into observable variables. Use it to track the AI capability timeline or place a single infrastructure company within the wider transmission chain.

• Anyone judging whether the AGI timeline and AI capability trends are still on track
• Anyone breaking down bottlenecks across AI compute, power, chip packaging, and data centers
• Anyone assessing which bottleneck or capability supports an AI company's valuation premium
• Anyone who wants to know which signals would weaken the AI-infrastructure thesis—and when to become more cautious

Boundaries: the trendline is not an oracle

"AGI by 2027" is the midpoint of a trend extrapolation, not a promise. It won't quote an exact AGI arrival date, won't time black swans, and won't recite real-time prices, moves, market caps, or valuations from memory.

When specific names come up, it says so plainly: this is extrapolation along the trendline — not a conclusion from Situational Awareness, and not investment advice. When there's no signal, it won't force a direction.

Leopold

Bring an AI compute, power, networking, or semiconductor question. The agent will map the orders of magnitude, the bottlenecks, and the evidence that could change the thesis.

Built from public materials to demonstrate a Leopold-inspired AI infrastructure framework. It does not represent Leopold Aschenbrenner or provide investment advice.

What evidence would change the expected timeline for AGI? Which parts of the AI infrastructure stack could capture the most value? When does the AI trade peak, and what warning signals should I watch?

Ask Leopold

Send