Amp Raises $1.3B to Build an Independent AI Grid — Pooled Compute for Frontier Labs That Can't Afford to Own It
Anjney Midha spent two years at Andreessen Horowitz allocating GPU access to AI startups through the Oxygen program — deciding which labs got the compute they needed to train competitive models and which did not. He watched independent teams overprovision for training peaks, waste capacity during inference troughs, and lose 30 to 40 percent of their allocated FLOPs to scheduling inefficiency.
Amp PBC, his new company, raised $1.3 billion to solve that problem at infrastructure scale.
The Problem Amp Is Targeting
Frontier AI workloads are structurally spiky. A team runs a large training job, then shifts to inference, then idles. Each phase has different compute requirements, and individual teams cannot smooth their own demand curves — they either overprovision and waste, or underprovision and block progress.
Hyperscalers solve this through scale. A cloud provider with millions of GPUs can absorb any individual team’s spikes. But hyperscaler access comes with dependency, pricing that prioritises the largest customers, and infrastructure tooling built for general compute rather than the specific failure modes of frontier AI: silent data corruption, topology-aware scheduling, fast checkpointing across hardware.
Amp’s answer is a shared grid specifically for independent frontier teams. When one member is in training mode and another is in deployment mode, their demand curves partially cancel. The collective utilisation rate improves. Each member gets burst access beyond their own baseload.
The Public Benefit Structure
Amp is structured as a public benefit corporation. The stated mission is maximising global frontier AI output — not maximising returns to Midha or the fund’s LPs. This matters because the grid’s value is collective: members that contribute their idle capacity when not training get access to others’ capacity when they are. A PBC structure is one way to make that commitment credible.
The founding grid partners include several of the leading independent research labs. Midha’s connections through the Oxygen program and his a16z board seats at Mistral, Black Forest Labs, and LMArena gave Amp access to the relationship layer that matters most in frontier compute: who trusts you with their training runs.
What $1.3 Billion Buys
At current market rates, $1.3 billion covers roughly 15,000 to 20,000 Nvidia H100-equivalent GPUs plus operating costs — power, cooling, network, and the infrastructure engineering team required to run them reliably for frontier workloads. Amp’s technical staff includes alumni from Google’s global AI infrastructure engineering group.
The investment round includes strategic participation from hyperscale cloud providers and semiconductor companies — parties that are technically Amp’s competitors but have reason to support infrastructure that increases total frontier AI activity, which eventually converts into cloud and chip spend.
The Utilisation Thesis
Midha’s argument, articulated on Amp’s public site, is that frontier AI output is being constrained not by a shortage of capable researchers but by compute inefficiency at the team level. A focused, talent-dense team of five people can now produce what required a hundred people in 2022. The constraint is not headcount. It is reliable compute access without the organisational cost of becoming large enough to justify owning it.
If that thesis is correct, the market is underinvested in the infrastructure layer between individual labs and hyperscale providers. Amp is the bet that the middle layer is durable — that even as clusters get larger and models get cheaper, the coordination problem for independent teams does not go away.
The counter-argument: GPU supply is not perennially scarce. If Nvidia and its foundry partners close the demand gap, Amp’s differentiation compresses. The grid’s other value — shared infrastructure tooling for frontier-specific failure modes — becomes the durable moat, or it does not.
Notable Detail
Amp appeared as a named investor in Anthropic’s Series G at $380 billion valuation — an early signal of which labs the grid is already serving. The company is now inviting mission-aligned organisations to apply for membership at grid@amppbc.com.