AMP launches $100B Grid Trust with Zurich Insurance to give startups access to pooled GPU compute

AMP Grid Trust

AMP, an AI infrastructure and venture firm headquartered at the Presidio in San Francisco, is launching a $100 billion compute program called the Grid Trust, anchored by Zurich Alternative Asset Management, the alternatives arm of Zurich Insurance Group with $440 billion in assets under management.

The premise is straightforward: startups that need GPU compute at scale are being priced out or turned away because they lack the purchasing power of hyperscalers. The Grid Trust pools capacity across customers to solve that problem structurally rather than one cluster at a time.

The grid is already live at nationalcompute.com, with over 60 commercial labs running on it. The hardware mix includes NVIDIA and AMD GPUs, with AMP aiming to get all major chip architectures onto the grid to give customers hardware flexibility. Public sector users — .edu, .gov, and .mil — get subsidized access: AMD MI355 chips at $3/hour and NVIDIA B200s at $4/hour. Commercial labs pay market rates, with MI355 access running around $12/hour on demand. AMP load-balances utilization across both tiers, using the commercial revenue to subsidize public sector access.

“We're kicking off a $100,000,000,000 program called the Grid Trust that's been anchored by Zurich. Zurich Asset Management has over 440,000,000,000 of assets under management, and they have anchored the Grid Trust program with a meaningful amount of capital that we are using to bring up compute capacity for startups all around the world. An AMD MI355 chip goes for $3 an hour for public sector users and $12 an hour for commercial customers.”

William Chew, Zurich's head of private equity, frames the investment around three things: the durability of compute demand regardless of which models win; the diversification benefits of a multi-tenant pool versus single-tenant buildouts; and the infrastructure-as-utility spending pattern that makes the asset class legible to institutional investors. He argues that while frontier labs and models cycle in and out, GPU consumption has grown consistently — the kind of observable, durable demand pattern Zurich looks for.

Negin Oliver, who spent eight years at AMD and helped assemble AMD's two-gigawatt partnerships with both OpenAI and Anthropic, is running the buildout. AMP is targeting two gigawatts of capacity, with Negin describing clear line of sight to capacity coming online over the next twelve months across multiple cloud providers and neo-clouds on the grid.

On demand forecasting — consistently one of the hardest operational problems for AI companies — AMP's answer is pooled visibility plus an in-house AI agent called Marshall, named after the postwar Marshall Plan. Grid members share their demand projections with AMP, Marshall ingests that data and produces forecasts, and AMP takes the bets on unexpected demand spikes. The logic is that with enough members, over-planners and under-planners smooth each other out. AMP is using reinforcement learning to improve Marshall over time, with the longer ambition of replacing what today requires 100-plus-person infrastructure and procurement teams at the largest labs.

The two-gigawatt target is ambitious and not yet delivered — the grid is live but the full capacity buildout is a multi-year effort. Whether pooled demand forecasting can actually absorb the kind of discontinuous demand spikes that accompany viral model releases or breakout product moments is an open question. But the commercial signal is real: several million dollars have already been spent on MI355 capacity at the $12/hour rate.