Physical Infrastructure As The Binding Constraint For Ai Scaling

Issue 104 Edition 2026-04-14 8 min read

Not accepted General

Sources: 1 • Confidence: Low • Updated: 2026-04-15 03:41

Key takeaways

A16Z invested in a company making physical power transformers because grid equipment innovation and manufacturability are needed, and transformer designs have changed little since early electrification.
Companies relying on legacy lock-ins for pricing will face strong pricing pressure and will need to anchor pricing to a more distinct value they provide.
AI will make cryptographic verification of personhood, identity, and content authenticity increasingly necessary because deepfakes and synthetic media will become indistinguishable.
A16Z's first fund was $300 million raised mainly from traditional US institutional LPs, while it recently raised $15 billion for four of seven funds from a more international and materially different LP base.
Electricity scarcity could entrench large AI companies that can secure power and GPUs or could push computation to the edge if models become small and efficient enough.

A16Z invested in a company making physical power transformers because grid equipment innovation and manufacturability are needed, and transformer designs have changed little since early electrification.
The United States lacks sufficient rare earth minerals, electricity, and manufacturing capacity to meet near-term AI-era infrastructure needs.
Electricity scarcity could entrench large AI companies that can secure power and GPUs or could push computation to the edge if models become small and efficient enough.
The United States is effectively out of electricity capacity right now, while AI token demand is rising very rapidly and buildout capacity is not keeping up.
AI supply-chain constraints will persist sequentially such that chips may become sufficient before electricity and memory, making downstream bottlenecks the limiting factors.

Companies relying on legacy lock-ins for pricing will face strong pricing pressure and will need to anchor pricing to a more distinct value they provide.
During AI-driven dislocation, CEOs should assess whether their business is strengthening or degenerating, and if customers have shifted spend away they may need deep cuts and a pivot.
Sufficient access to GPUs plus good data can dramatically compress software development timelines, weakening the classic idea that money cannot buy you out of a software deficit.
Traditional SaaS lock-ins are weakening because code and data are easier to replicate and AI agents reduce the importance of human-facing interfaces.
Navan is defensible due to hard-to-replicate supplier relationships and enterprise sales channels.

AI will make cryptographic verification of personhood, identity, and content authenticity increasingly necessary because deepfakes and synthetic media will become indistinguishable.
Blockchains are a preferable trust anchor for authenticity and provenance because they avoid reliance on any single corporation or government as arbiter of truth.
If AIs become economic actors, they will need native internet money because traditional merchant and payment systems are poorly suited to non-human entities.
Large-scale government payments without strong identity and address infrastructure are highly vulnerable to fraud, and stimulus-program theft was around $450 billion.

A16Z's first fund was $300 million raised mainly from traditional US institutional LPs, while it recently raised $15 billion for four of seven funds from a more international and materially different LP base.
A16Z invested in a company making physical power transformers because grid equipment innovation and manufacturability are needed, and transformer designs have changed little since early electrification.
A reason for raising larger funds is the belief that the US must rapidly fund rebuilding AI-era economic infrastructure and that the required investment will be very large.

Electricity scarcity could entrench large AI companies that can secure power and GPUs or could push computation to the edge if models become small and efficient enough.
AI markets may consolidate into a small number of gigantic companies that own most of the value.
Frontier models may become utility-like and could be regulated or nationalized, enabling many companies to build on standardized AI infrastructure.

What specific, measured evidence supports the claim that the US is currently constrained by electricity capacity for AI data centers (by region, timeframe, and magnitude)?
Are memory constraints (and which kinds) actually becoming more limiting than chips, and on what timeline relative to power constraints?
Do SaaS vendors relying on legacy switching costs show measurable increases in churn, renewal discounting, or net revenue retention declines attributable to AI-driven substitutability?
To what extent do AI agents actually bypass human-facing interfaces in production workflows, reducing UI lock-in as a moat?
What concrete standards or deployments are emerging for cryptographic content provenance and personhood verification, and what adoption thresholds (enterprise or public-sector) are being met?

Grid hardware and power availability may become primary constraints on AI scaling, shifting value toward manufacturers and innovators of transformers and related equipment rather than purely software layers.
AI could erode classic SaaS lock-in and force pricing anchored to distinct value, shifting competition toward distribution, relationships, and outcomes rather than UI and switching costs.
Demand for cryptographic identity and content provenance infrastructure could rise as synthetic media becomes indistinguishable, creating new trust layers used by enterprises and public-sector payment flows.

Public, region-specific evidence of power or interconnection constraints for data centers, including long lead times for transformers and grid upgrades and capex plans prioritizing power delivery.
SaaS cohort metrics show higher churn, greater renewal discounting, or declining net revenue retention coincident with AI-enabled substitution, alongside vendor messaging shifting to outcome-based pricing.
Concrete standards and deployments for content provenance and personhood verification gain measurable adoption, especially in enterprise workflows or government disbursement systems with fraud reduction goals.

Data indicates electricity capacity is not binding for AI buildouts, or constraints quickly ease without material grid hardware bottlenecks, reducing the importance of transformer-focused innovation.
SaaS incumbents sustain pricing power and retention without increased discounting or churn, and AI does not materially reduce UI or migration friction in production workflows.
Provenance and identity initiatives fail to reach meaningful adoption thresholds, or non-cryptographic approaches dominate without requiring new trust infrastructure spending.