ExpertsFMChina is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer
Dwarkesh Podcast · August 15, 2025
5 Selected Lessons:
reasoning_pattern · 2 heuristic · 1 prediction · 1 priority · 1
Minimalist first-principles design: delete everything not absolutely needed
Source
Dwarkesh Podcast — China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer (August 15, 2025) · at 59:30 ▸
When this applies
when designing any system and tempted to include conventional supporting infrastructure by default
The takeaway
Default engineering/infrastructure stacks are usually larger than physically necessary; the discipline of asking 'do we absolutely need this component' at every layer reveals a much leaner achievable design.
Why it works
By stripping a design down to only the truly necessary components (e.g., skip grid, transformers, even DC-DC converters if a buck converter/relay suffices) you reveal that most 'standard' infrastructure exists for legacy/regulatory reasons rather than physical necessity, exposing a far simpler and cheaper achievable architecture.
One level up
True first-principles design requires actively hunting for and removing components justified only by convention, not necessity.
Where it stops holding
This stops working when the 'unnecessary' component actually provides essential safety, regulatory compliance, or resilience margin the minimal design lacks.
Detail
hierarchy:
- Necessary: silicon array (power) + silicon compute
- Needed on Earth but not in space: battery for storage
- Needed only if off-grid Earth system lacks battery/solar colocation: interconnects/relays
- Not actually needed: grid, ISOs, transformers, full DC-DC conversion stack
rationale: The first thing Elon will always ask is, like, delete anything you don't absolutely need — this yields a theoretical minimal AI power+compute system of just two slabs of silicon.
In their words
the first thing that Elon will always ask is, like, delete anything you don't absolutely need. So what you actually need is a big slab of relatively cheap silicon to make the power and then a small slab of relatively expensive silicon to do the thinking.
'Write the check' skepticism toward claimed industry limits
Source
Dwarkesh Podcast — China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer (August 15, 2025) · at 20:49 ▸
When this applies
whenever an industry insider says a capability expansion 'can't' happen
The takeaway
Treat 'we can't do it' statements from incumbents as really meaning 'no one has yet funded/derisked it,' and test that by looking at what happened once capital was actually committed.
Why it works
Claims that an industry 'can't' scale usually reflect unresolved financing/risk allocation, not a true physical or technical ceiling; once someone commits capital (writes the check) and absorbs the risk, production ramps.
One level up
Apparent capability ceilings are frequently financing and risk-allocation problems in disguise, not technical ones.
Where it stops holding
This stops working when the limiting factor truly is a hard physical/technical constraint (e.g., a genuine materials or physics ceiling) rather than financing risk.
Detail
approach: When an incumbent says 'no way,' translate it as 'write me a check' and look for whether that capital eventually arrived and resolved the bottleneck.
signals:
- twenty-year payback horizons
- uncertainty about demand persistence
- competitor or geopolitical risk cited as reason to not invest
rationale: The HBM case shows once demand was proven and backed with capital/commitment, manufacturers that previously refused to expand did so (Samsung building in the US with xAI).
In their words
Well, so when someone says we can't do it, we won't do it, no way, no how
New AI-native job categories get priced against AI competitors, not against human wages
Source
Dwarkesh Podcast — China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer (August 15, 2025) · at 55:22 ▸
When this applies
when a new technology creates a role where it doesn't compete against the old incumbent but against other versions of itself
The takeaway
Don't assume a new capability will be priced near the wage of the human role it most resembles — if it creates a genuinely new niche, its price will be set by competition among providers of that capability, converging toward their marginal cost.
Why it works
If a new role emerges that's well-suited to a new technology (rather than an existing human job being automated), pricing isn't anchored to the previous human wage but gets competed down toward the new technology's own marginal production cost, because the competition is between providers of the new technology (labs), not between new tech and old labor.
One level up
Pricing in a market is set by the actual competitive set, not by analogy to a superficially similar but structurally different market.
Where it stops holding
This stops working when the new role is still in direct substitution competition with human labor (not a novel niche), in which case anchoring to wages remains valid.
Detail
approach: For a new AI-native role, ask who AI is actually competing against (other AI labs, not humans) and expect price to fall toward marginal token cost rather than stay anchored near the comparable human wage.
signals:
- role didn't previously exist for humans in that exact form
- AI providers compete against each other for that workload
- cost converges toward compute marginal cost, not toward prior wage levels
rationale: A $200k/year human researcher role being matched by AI doesn't mean the AI service is priced at $200k — it could be priced far higher in value-capture terms or far lower in marginal-cost terms, depending on dynamics.
In their words
it'll be competing against the other labs. Right? And so you'd actually see, like, the cost push down to, you know, a small multiple of whatever the marginal production cost of those tokens is.
Bragging about infrastructure prowess signals bad capital allocation
Source
Dwarkesh Podcast — China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer (August 15, 2025) · at 0:00 ▸
When this applies
when an organization proudly points to heavy investment in a capability as proof of superiority
The takeaway
Visible scale of investment is not evidence of good capital allocation; ask what return that capital is actually generating.
Why it works
Building large amounts of capital-intensive infrastructure that doesn't generate proportional returns indicates capital is being misallocated rather than well-deployed.
One level up
Judge resource allocation by returns generated, not by the scale or visibility of the effort.
Where it stops holding
This stops working when the large build-out is genuinely the cheapest path to a necessary strategic capability rather than a vanity project.
Detail
rule: If a country or company holds up massive infrastructure buildout as a point of pride, question whether that's really evidence of good capital allocation.
threshold: n/a - qualitative judgment
rationale: Capital devoted to projects with questionable ROI, even at huge scale, is a sign of inefficiency, not strength.
In their words
Like, that is just a sign that you're really bad at capital allocation. Like, why would you devote in 2025 so much industrial effort and money?
Most new data centers built from 2027 onward will be mostly solar
Source
Dwarkesh Podcast — China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer (August 15, 2025) · at 24:58 ▸
When this applies
forecasting technology adoption inflection points based on current planning conversations with insiders
The takeaway
Infer near-term technology adoption inflection points from what's currently being planned/procured, since multi-year lead times mean today's decisions determine tomorrow's deployment mix.
Why it works
Because planning for 2027 groundbreaking is happening now, and multiple hyperscalers are already calling Casey about 90%+ solar microgrid designs, the shift to majority-solar new builds is already locked in by current planning decisions.
One level up
The present planning pipeline is a leading indicator of near-future infrastructure composition.
Where it stops holding
This stops working if hyperscalers hit unforeseen constraints (transformer/grid bottlenecks, permitting delays, land acquisition issues) that force them to default to gas/grid instead.
Detail
rule: n/a - prediction
threshold: By 2027, majority of new data center groundbreakings will be mostly solar (90%+ solar / 10% gas or better)
rationale: Based on direct insider calls Casey receives about near-term (next-few-years) plans at 5GW scale.
In their words
So I think within a few years, we'll probably see that, like, the majority of new DCs that are going in will be mostly solar.