随着 CoreWeave 股价暴跌，拖累了 AI 板块的大部分股票，投资者面临一个简单的问题：对冲 AI 风险的最佳方式是什么？一位前 Citadel 量化研究员可能给出了答案。
As CoreWeave shares plummet, taking much of the AI complex with it, investors face a simple question: what's the best way to hedge AI risk? An ex-Citadel quant might have the answer.
Jasper Zhang（张杰，音译）曾任 Citadel Securities 量化研究员，现领导 Hyperbolic。这是一家 50 人的金融科技初创公司，正在构建一个可交易 GPU 算力的现货市场，该工具将允许交易者对冲 AI 需求下降的风险。该平台允许投资者买卖 GPU 集群，即为 AI 算力提供动力的单元。这是通过 Hyperbolic 的 Forge 软件实现的，该软件虚拟登录全球各地的集群并进行技术调整，使配置不匹配的集群能更好地协同工作，从而让交易者不仅能买卖标准化的算力单元，还能为自己的目的控制底层 GPU。
Jasper Zhang is an ex-Citadel Securities quantitative researcher who now leads Hyperbolic. It's a 50-person financial technology startup building a spot marketplace for tradable GPU computing power, a tool that would allow traders to hedge against a decline in AI demand. The platform allows investors to buy and sell GPU clusters, the units that power AI compute. This is done through Hyperbolic's Forge software, which virtually logs into clusters worldwide and makes technical tweaks so that mismatched clusters work better together, allowing traders to not just buy and sell a standardized unit of compute, but also control the underlying GPUs for their own purposes.
张杰表示，这相当于身处数据中心本身，并将该技术视为“实物”交割 GPU 期货的第一块基石，此类期货在交割时将赋予合约持有者对已知特性集群的完全控制权。换句话说，这将允许投资者通过交易硬件本身，以一种新方式对冲 AI 需求下降的风险。
Zhang says it's as good as being in the data center itself, and sees the technology as the first building block of "physically" settled GPU futures, which, upon settlement, would offer contract owners complete control over a cluster with known characteristics. In other words, it would allow investors a new way to hedge against a decline in AI demand by trading the hardware itself.
目前，GPU 架构因组装方式不同而差异巨大，导致输出水平各异。每个数据中心都不一样。网络、内存和安全都会影响 GPU 性能。同时，融资协议往往禁止建设方共享价格，这意味着基于调查的指数可能只能平均建设成本，而遗漏异常开支。在期货领域，这被称为“基差风险”。这是做市商需要解决的问题，也是张杰希望 Hyperbolic 能填补的空白。
Currently, GPU architecture varies wildly depending on how they are assembled, leading to various levels of output. Every data center is different. Networking, memory, and security all affect GPU performance. At the same time, financing agreements often prevent builders from sharing prices, meaning a survey-based index may average setup costs while missing outlier expenses. In the futures world, that's called "basis risk." It's a problem for market-makers to solve, and a void Zhang hopes Hyperbolic can fill.
现年 31 岁的张杰带着几枚中国数学奖牌来到美国，仅用两年便在加州大学伯克利分校完成数学博士学位，随后入职 Citadel Securities 构建交易模型。
The 31-year-old Zhang came to America with several Chinese math medals and finished a math doctorate at U.C. Berkeley in just two years before taking a job at Citadel Securities building trading models.
Hyperbolic的财务负责人金洪灿（Hong Chan Kim）是一位对冲基金资深人士，曾任职于Point72、Balyasny Asset Management和Citadel LLC等对冲基金。
Hyperbolic's head of finance, Hong Chan Kim is a hedge fund veteran, having worked at Point72, Balyasny Asset Management and Citadel LLC, the hedge fund.
自2022年成立以来，Hyperbolic已筹集2000万美元资金，目前正处于另一轮融资之中，并预计到年底实现1.5亿美元的年化营收运行率。截至目前，该公司表示拥有约25万用户，并估计有价值约2亿美元的年度GPU算力接入其平台。
Hyperbolic has raised $20 million since its launch in 2022, is in the midst of another funding round, and expects a $150 million annualized run rate by year-end. As of today, the company says it has about 250,000 users and an estimated $200 million worth of annual GPU compute connected to its platform.
其他主要参与者也看到了AI对冲工具市场中类似的机遇。CME集团计划推出一种现金结算期货产品，该产品挂钩由Silicon Data维护的GPU定价指数；Silicon Data是一家由DRW支持的指数提供商，DRW是芝加哥最大的自营交易公司之一。DRW创始人兼首席执行官Don Wilson在一次电话会议上表示：“大量的需求和资金正涌入这一生态系统，因此存在大量风险。现金期货是一个很好的第一个基石。有许多电力期货是现金结算的，石油期货也是现金结算的，所以我认为你可以用任一方式启动这些产品。”
Other big players see a similar opportunity in the marketplace for AI hedging tools. CME Group plans to offer a cash-settled futures product that is tied to GPU pricing maintained by Silicon Data, an index-provider backed by DRW, one of the biggest prop-trading firms in Chicago. "There's a ton of demand and capital flowing into this ecosystem so there's a lot of risk," Don Wilson, founder and CEO of DRW, said in a call. "Cash futures are a great first building block. There are plenty of electricity futures that are cash-settled, oil futures that are cash-settled, so I think you can start these things either way."
该CME产品目前正接受CFTC（美国商品期货交易委员会）审查，但该委员会最近将审查期限延长至11月9日，这很可能也拖延了洲际交易所（Intercontinental Exchange）和数据提供商Ornn提出的另一份期货合约提案，后者由Andreessen Horowitz的a16z加密货币基金支持。Hyperbolic的Kim表示：“我们猜测，当现金结算期货推出时，它们不会反映现实中正在发生的情况。实物使用量才会将现实与衍生品之间的联系系牢。”
The CME product is currently under review with the CFTC, but the commission recently extended that review period to November 9th, likely also tying up another futures contract proposed by Intercontinental Exchange and data-provider Ornn, backed by Andreessen Horowitz's a16z crypto fund. "Our guess is when the cash-settled futures come out they're not going to be reflective of what's going on in reality," says Hyperbolic's Kim. "Physical usage is what will tether that connection between the reality and the derivatives."
更大的问题是，现金期货对现实世界风险的对冲并不完美。如果AI需求撞上墙壁，像CoreWeave这样的AI租赁中心可能会手握数十亿美元的AI硬件，却几乎没有变现渠道。“指数是一个标准化的全球数字，”Silicon Data首席执行官兼DRW支持的Compute Exchange首席执行官Carmen Li表示，Compute Exchange是一个买卖GPU的市场，也在为实物交割期货奠定基础。
The bigger issue is that cash futures are an imperfect hedge for real-world risk. If AI demand hits a wall, AI rent-a-centers like CoreWeave could end up sitting on billions of dollars in AI hardware with little to no way to cash out. "Indices are a normalized global number," said Carmen Li, CEO of Silicon data and DRW-backed Compute Exchange, a marketplace for buying and selling GPUs that is also laying groundwork for physically-settled futures.
“你在财务上得到了对冲，但如果你是在寻找采购渠道，那没用。”Zhang主张，AI产业最需要的——特别是那些没有资本或信用去承担大规模硬件采购风险的小型初创公司——是一条通往实物硬件的虚拟管道。
"You get financially hedged, but if you're looking for procurement, that's not going to help you." What the AI complex needs most – in particular smaller startups that don't have the capital or credit to risk making big hardware purchases – is a virtual pipeline to physical hardware, contends Zhang.
“如果供应商在放缓期有更多方式将可用的 GPU 产能变现，尤其是在长期合同利润微薄的情况下，CoreWeave 的债务状况就不会那么令人担忧，”张某说。“现金结算期货有助于应对价格下跌，但它们本身无法让闲置的 GPU 发挥作用。”
"CoreWeave's debt situation would be less concerning if providers had more ways to monetize available GPU capacity during a slowdown, especially when long-term contracts carry thin margins," said Zhang. "Cash-settled futures can help falling prices, but they do not by themselves put idle GPUs to work."