Research · November 6, 2026

Hyperbolic's $30M Series A1 — Making GPU Compute Liquid, Not Just Available

Hyperbolic raised $30 million in a Series A1 round led by Walden Catalyst to expand its Forge control plane and build a Global Compute Grid — a marketplace where eligible GPU capacity can be bought, sold, and resold across providers.

What happened

On October 27, 2026, Hyperbolic, a San Francisco-based AI infrastructure company, announced a $30 million Series A1 round led by Walden Catalyst, the venture firm founded by Intel CEO Lip-Bu Tan and former Samsung Electronics president and chief strategy officer Young Sohn. Dell Technologies Capital, General Catalyst, Vista Equity Partners, Cambium Capital, and Salience Capital also participated, along with other strategic investors. Reporting on the round puts Hyperbolic's total funding at $50 million; the company's own announcement discloses only the $30 million round, so the cumulative figure is attributed to reporting here rather than stated as a company figure.

Hyperbolic's premise is that the AI compute shortage is partly a coordination failure, not just a capacity shortage. By the company's count, the GPU cloud market now spans more than 300 providers with different hardware, networking, software, and operational standards. AI teams face complex procurement, rigid contracts, and engineering overhead, and often pay for capacity that sits idle between workloads — while providers struggle to turn hardware, power, and data center investments into consistent revenue. Hyperbolic's answer is a layer above the providers: software that makes fragmented capacity behave like one pool.

Forge, the control plane underneath the pitch

The core product is Forge, a unified control plane that standardizes heterogeneous bare-metal GPU clusters into validated infrastructure, with a common dashboard for provisioning, virtualization, networking, and operations across providers. Managed Kubernetes and inference services sit on top, and a unified API gives teams access to models hosted by Hyperbolic and by external providers. The new money expands multi-cloud orchestration — provisioning compute and placing workloads, including inference, across providers based on cost, performance, and availability — with planned additions of Slurm support, automated scaling, and unified data services.

The founder's background explains the second half of the pitch. Founder and CEO Jasper Zhang completed a mathematics PhD at UC Berkeley in two years and then worked in market making at Citadel Securities, and the broader team draws on Google Cloud, Lambda, Lightning AI, VMware, Supermicro, Citadel, and Point72. That is the pedigree behind the round's most ambitious claim: the Global Compute Grid, a planned marketplace where suppliers and customers can buy, sell, and resell eligible GPU capacity, with spot instances, dynamic pricing, and market-making strategies to match supply with demand — and exploration of ways for customers and suppliers to hedge compute price risk. The stated vision is compute that works like electricity: draw capacity as needed, scale up or down, and sell eligible unused capacity back to the grid.

Walden's Lip-Bu Tan framed the investment in utilization terms: "The demand for AI compute will continue to grow, but adding more hardware is only part of the answer. The value of that investment depends on how effectively the capacity is put to work and made available to the teams that need it." Zhang's version goes further, arguing that AI agents — which "are not going to wine and dine with 20 different neocloud sales or sign contracts lasting three years or more" — will become the largest consumers of compute, and will need a programmable way to discover, acquire, and run capacity within defined budgets and policies.

The demand signal the company leads with is adoption: more than 250,000 AI builders use Hyperbolic, including teams at Modal, Higgsfield, Daytona, HeyGen, Voio, RadixArk, and Pika. Higgsfield's account is the most concrete — it says it obtained H100 capacity quickly for production inference, then converted to a reserved contract at a rate 35% cheaper than on-demand and 20% cheaper than anything else it was seeing in the market. Those are customer-reported figures in the company's own announcement, not independently verified results, but they illustrate the arbitrage Hyperbolic is selling: the same GPU hour, bought at a better time and on better terms.

Why this is a different bet from the compute sellers

This is worth distinguishing from the other compute bet in this queue. General Compute (draft 27) sells GPU compute capacity itself — it owns or controls the machines. Hyperbolic owns no capacity; it sells orchestration and, eventually, liquidity across other people's machines. If GPU supply keeps expanding while utilization stays uneven, the scarce asset may not be the GPU at all, but the layer that routes a workload to the cheapest available hour and finds a buyer for the hour that would otherwise sit idle. Hyperbolic's round is a bet that this layer is a market, not just a tool — and that a market maker's instincts, applied to compute, are the right ones for building it.

What this does not show

The announcement discloses no valuation, no revenue or ARR figure, no pricing for Forge or the planned grid, and no launch date, eligible-capacity rules, or fee structure for the Global Compute Grid — the marketplace is a plan, not a live product. The $50 million cumulative funding figure comes from reporting, not from the company's announcement, and the prior $12 million Series A in December 2024 is likewise reported rather than confirmed in the announcement. The 300-provider count and the 250,000-builder figure are company-reported, and the Higgsfield savings percentages are a customer account published by the company, not an audited result.

What would turn this round from an orchestration story into a compute-market story is evidence from the grid itself: whether providers actually list eligible capacity, whether spot pricing and resale clear at meaningful volume, and whether utilization gains survive contact with real contracts, interconnect limits, and the heterogeneity Forge claims to standardize away. Until then, Hyperbolic's case rests on a checkable intermediate claim — that one dashboard can already run workloads across many providers — and a much larger unproven one, that GPU hours can trade like electricity.

Sources