DeepTech Returns: Capital Rotates to Atoms

DeepTech Returns: Capital Rotates to Atoms

The largest new funding rounds of 2026 are not pure-software. They are energy, semiconductors, advanced manufacturing, defense and climate hardware — "atoms over bits". The driver is that the binding constraint on AI is now physical: power, fab capacity, HBM supply and cooling, not model architecture.

The constraint stack

What actually limits AI buildout in 2026 (top = tightest)
  1. Grid interconnection + firm power   <- 24-48 month queues
  2. HBM4 / advanced packaging capacity
  3. Datacenter shell + cooling
  4. Leading-edge wafer starts
  5. Model quality                        <- no longer the bottleneck

Decision matrix

Sector	Why capital is moving	Risk
Firm power (SMR, geothermal, grid)	AI load growth outpaces utility planning	Regulatory + build timelines measured in years
Advanced packaging / HBM	Single supplier concentration, structural shortage	Cyclical; capacity gluts historically follow shortages
Domestic advanced manufacturing	Policy tailwind + defense demand	Capex-heavy, long payback
Climate hardware	Mandates + corporate procurement	Policy-dependent demand

Playbook

1. If you buy compute, your procurement risk is power and packaging lead time — contract for both 18+ months out.
2. DeepTech diligence is different: unit economics at scale, not ARR growth. Model the nth-of-a-kind cost, not the first.
3. Expect longer holds. These are 7-10 year outcomes, not 4.