AMP PBC Raises $1.3B for GPU Utility Grid to Democratize AI Compute
In brief
- AMP PBC pools underutilized GPUs from independent labs and data centers into a shared compute grid
- Startup raised $1.3 billion from Andreessen Horowitz, Y Combinator, and other backers
- GPU rental rates doubled since January 2026, pricing out smaller AI teams
- AMP operates as public benefit corporation with compute grid and venture arm
- Model competes with blockchain-based alternatives like Render Network and Akash
The GPU rental crunch
Cloud GPU rental prices have surged, with hourly rates for B200 units reaching $4.89. Those rates have roughly doubled since January 2026, squeezing smaller AI teams that can't match the purchasing power of Google, Microsoft, and Meta.
That's the structural problem Midha identified after his tenure at the venture firm. The solution: pool idle capacity from data centers and independent labs, then let AI teams draw from the shared reserve. Data center operators with idle GPUs contribute capacity to the network and can sell excess capacity for profit.
The $1.3 billion bet
AMP PBC has $1.3 billion in funding commitments from backers including Andreessen Horowitz and Y Combinator. The company operates as a public benefit corporation, a legal structure that lets the company pursue social objectives alongside profit.
The firm has two arms: a compute grid and a venture operation. The venture side funds AI teams building on top of the grid, aligning incentives across the network. Midha has been making the rounds to explain this vision, including a May 5 appearance on the TBPN podcast and a June 2 CNBC segment.
Centralized versus decentralized
The competitive landscape is crowded. CoreWeave has built a massive GPU cloud business, Lambda Labs offers on-demand GPU clusters, and Together AI focuses on distributed inference. On the blockchain side, Render Network, Akash, and io.net all pursue variations of this idea using blockchain-based coordination layers.
AMP's differentiator is its centralized approach. AMP takes a centralized model, which may sacrifice the ideological purity of decentralization but could gain advantages in coordination speed and operational simplicity. Whether that trade-off pays off depends on execution and adoption rates among data center operators willing to join the network.
