Anj Midha's AMP raises $1.3B to build a shared power grid for AI compute

What's the deal? AMPDealroom has a profile for this one. Try Dealroom →, a Menlo Park-based AI infrastructure venture founded by Anjney MidhaDealroom has a profile for this one. Try Dealroom → — former general partner at Andreessen Horowitz and an early personal investor in AnthropicDealroom has a profile for this one. Try Dealroom → — has raised $1.3B for its first fund.

Investors include Andreessen Horowitz, Y Combinator, and several cloud computing providers. AMP has already made a $300M investment in Anthropic as part of its $30B funding round in February 2026. The firm is also raising capital for a second venture fund and working on a separate $10B infrastructure credit fund.

AMP's core thesis is that AI compute is being hoarded by a small number of wealthy organisations, leaving startups, universities, and researchers unable to access the infrastructure they need.

Rather than building or owning data centres, AMP buys compute capacity from data centre operators and pools it into a shared resource — a coalition that AI developers can join to access servers they could not otherwise afford or secure.

Several startups have already joined, including Periodic Labs and ElevenLabs. AMP's name comes from "ampere," the unit of electrical current — a deliberate reference to its grid analogy.

Why now? AI compute demand is highly unpredictable — training runs create sudden large spikes followed by quieter periods, making it difficult for developers to reserve exactly the right amount of capacity. The result is a market where developers either over-reserve and waste money or under-reserve and can't scale when they need to.

Meanwhile, the largest AI labs and cloud providers have locked up vast quantities of GPU capacity through long-term contracts, leaving everyone else competing for scraps.

Midha developed the concept at Andreessen Horowitz, where he ran Oxygen — an internal programme that pooled Nvidia chips for the firm's portfolio companies. He spun it out as AMP because he believed the problem was too large and too important to solve within the confines of a single VC firm's portfolio.

What could go wrong? AMP's model requires buy-in from both sides of a two-sided market simultaneously — data centre operators willing to sell capacity to a neutral intermediary, and AI developers willing to pool their compute needs through a shared system rather than securing their own infrastructure. AI firms currently view compute access as a strategic advantage and may be reluctant to share or pool resources even if the economics are better.

AMP also sits in an unusual position — part VC firm, part infrastructure operator, part collective bargaining body. That hybrid model has no obvious precedent and could face regulatory, commercial, or structural challenges that a more conventional business model would not.

The signal: AMP is one of the most ambitious attempts yet to restructure how AI infrastructure is allocated. The electricity grid analogy is not just marketing — it reflects a genuine belief that AI compute will eventually be commoditised and socialised the way power was in the early 20th century. If that thesis is right, AMP could become foundational infrastructure for the AI era. If it's wrong — if compute remains a strategic asset that large players refuse to share — the model may struggle to achieve the scale needed to matter.

Midha's ties to Anthropic, a16z, and the broader frontier AI ecosystem give AMP an unusually credible starting position for a firm making such an audacious bet.

Sources:
AMP
The Information
The Information
The Information
New York Times
Startup Fortune

Image credit:
Andreessen Horowitz

J.V.