On 26 October, Jasper Zhang told CNBC that about 250,000 developers use his platform, and that roughly $200 million of GPU capacity connects to its software each year. The next morning, Hyperbolic announced $30 million in Series A-1 funding. Walden Catalyst led the round, bringing total capital to $50 million.
Zhang is 31. He finished a Berkeley PhD in two years, then ran market-making models at Citadel Securities. He now applies the same logic to Nvidia GPUs.
The pitch fits in one line. GPUs are financed for years, while AI workloads run for weeks. That mismatch between contract length and workload length is the thing Hyperbolic wants to price.
Three hundred clouds, one contract problem
GPU capacity is sold by more than 300 providers, each with its own hardware mix, network topology and minimum commit. A team that needs 128 H100s for six weeks negotiates with several of them, then pays for headroom it may not use. Hyperbolic works with about 30 neoclouds, including WhiteFiber and Vector Core Compute, and claims to put their idle capacity behind one interface.
Forge, launched in June, standardizes bare-metal clusters into validated infrastructure with a common provisioning and networking layer. Flux, launched on 21 October, adds model serving and token metering on top. A customer can rent a cluster, run inference through an OpenAI-compatible endpoint, and move capacity between workloads as demand shifts. Customers named in the announcement include Modal, Daytona, HeyGen and Pika.
The build order
Hyperbolic is assembling a market in stages, and the sequence matters. Orchestration comes first, because a market needs a standard unit to trade. The grid comes second. Derivatives come last, if they come at all.
CHRONICLE: Hyperbolic, 2022–2026
────────────────────────────
2022 · 🧪 Founded by Jasper Zhang after Citadel Securities
2024 · 🚀 Serverless inference opens to outside developers
2026 · 🧰 Forge standardizes bare-metal GPU clusters
2026 · 🧠 Flux moves inference onto the same control plane
◉ NOW · 💰 $30M Series A-1; $50M raised to date
🔥 NEXT · 🌐 Global Compute Grid with spot pricing and resale
Sources: Hyperbolic announcements, June–October 2026
The $200 million of connected capacity is the inventory; a $150 million annual run-rate target by year-end is the test of its take rate.
A compute futures market is a later product, and it depends on standardization the company has not finished.
From brokerage to market-making
The funding is earmarked for Forge, multi-cloud orchestration, and a Global Compute Grid where suppliers and customers can buy, sell and resell eligible GPU capacity. Planned features include spot instances, dynamic pricing and market-making strategies. Hyperbolic also says it is exploring ways for buyers and suppliers to hedge compute price risk.
The ambition is explicit. Zhang describes compute as something customers should draw like electricity, then sell back when they do not need it. The company already runs the brokerage layer: it matches workloads to capacity across providers. Market-making is a different job, because it requires the platform to hold inventory, quote two-sided prices, and absorb the risk of being wrong.
Hyperbolic says it is already breaking even and expects to reach 50 employees by the end of 2026. Those numbers are unusual for a company raising a Series A-1, and they explain the investor list. Angels include executives from Citadel Securities and the New York Stock Exchange, alongside GPU cloud operators WhiteFiber and Vector Core Compute. The money is not only betting on software margins. It is betting that compute becomes a financial asset with a yield curve.
One customer example shows where the pricing pressure sits. RadixArk's CTO said Hyperbolic converted it from on-demand H100 capacity to a reserved contract at a rate 35% below on-demand and 20% below other quotes it had seen. A resale market would compress that spread further, which helps buyers and squeezes providers who bought hardware on three-year assumptions.
A deliverable unit that means the same thing at two different providers.
Settlement and credit terms that survive a counterparty failure.
A reference price that buyers, lenders and auditors all accept.
The hard part is not the exchange
Building a matching engine is routine. Defining what gets delivered is not. Two H100 clusters can differ in interconnect, storage, cooling headroom and uptime history. A futures contract needs a specification that survives a hardware generation change, and the cadence has already moved from H100 to B200 to GB300 in about three years.
Power has become the binding constraint on the underlying asset. On 20 October, Microsoft walked away from a 50-megawatt site in Loughton, Essex, where Nscale planned to install 23,040 Nvidia GB300 chips. Grid connection queues in Britain can stretch to 13 years. A provider that cannot energize a site cannot deliver into a contract, and a market that cannot guarantee delivery is a directory with better branding.
Financing risk compounds the delivery risk. Neoclouds borrow against GPU fleets whose resale value depends on the next model generation. Hyperbolic's answer is liquidity: if unused capacity can be resold, lenders get a secondary market for collateral. That argument works only when the platform can clear trades at a price both sides trust.
The comparison with inference providers is instructive. In October, Fireworks AI was reported to be seeking a $40 billion valuation on a revenue figure last updated in July, as we wrote in October. Capital is already pricing AI infrastructure on forward utilization. A compute exchange would make those assumptions visible, and mark them daily.
What a price curve changes
If Hyperbolic succeeds, three things follow. Capex underwriting moves from nameplate capacity to realized utilization, which favors operators with diversified buyers. Provider consolidation accelerates, because small neoclouds gain a resale channel but lose pricing power. And procurement becomes programmable, which is the part Zhang is really selling.
His argument about agents is the strongest piece of the thesis. Agents will not negotiate twenty separate contracts. They will call an API, state a budget and a deadline, and expect capacity to appear. A platform that meters tokens and rents clusters can serve that request today. A platform that also quotes a forward price can serve it tomorrow.
The $50 million buys time to prove the grid can settle a trade. Hyperbolic has the inventory, the software and a founder who understands spreads. What it does not yet have is a standard contract that two strangers will honor when the hardware underneath changes. Until that exists, the futures market is a slide, not a product.