CME to Launch GPU Futures as Compute Market Financializes

算力市场迈向金融化：CME将推出GPU期货 Compute Exchange与Silicon Data正与CME Group合作推出标准化GPU期货，利用性能归一化解决芯片产出高达38%的差异问题。 本期节目探讨了GPU算力的金融化进程，以及将稀缺算力转化为可交易大宗商品的努力。Carmen Li详细介绍了Silicon Data如何接入庞大的每日定价数据集来构建基准指数，以解决芯片性能差异问题，而Compute Exchange则作为经过验证的硬件合约的现货市场运作。她分析了CME算力期货的机制、跨芯片对冲、二手硬件折旧曲线，以及A100、H100和B200等不同代际芯片之间的市场波动。讨论最终展现了长期远期合约与物理基础设施瓶颈如何稳定AI算力的底层经济机制。 Silicon Data与Compute Exchange正在构建基础金融基础设施——包括现货市场、定价指数以及由CME支持的衍生品——以将GPU算力转化为可交易的大宗商品。 (0:00) 算力市场的买家涵盖AI初创公司、传统企业和推理托管商，其中许多正从波动剧烈的按需价格（每小时3至9美元）转向远期和预留合约，以确保利润率的可预测性。 (3:52) 市场参与者的行为类似于传统大宗商品交易者：硬件所有者和新型云（NeoClouds）天然做多算力并做空期货以锁定收入，而AI开发者则天然做空算力并买入期货以对冲成本飙升风险。 (3:52) 受芯片体质、内存带宽及系统配置影响，相同型号的GPU存在高达38%的性能差异——这种被称为“GPU抽奖”（GPU Lottery）的现象需要独立测试与标准化基准评测。 (9:07) 为构建标准化定价指数，Silicon Data每日处理来自100多个数据源的逾15万条交易价格，维持日波动率在20%至30%之间的归一化指数。 (9:07)

Carmen Li, CEO of Silicon Data and Compute Exchange, is developing index benchmarks and derivatives infrastructure to transform GPU compute into a standardized, hedgeable commodity. Backed by DRW founder Don Wilson and partnering with CME Group, the initiative aims to provide pricing transparency and risk management tools for AI developers and cloud providers. Dual Infrastructure for Spot and Derivatives Trading (3:52) To establish compute as a tradable commodity, Li operates two entities with distinct mandates. Silicon Data, founded in April 2024, constructs benchmark pricing indices by tracking over 8 million pricing points from approximately 200 providers on the Bloomberg Terminal. Compute Exchange functions as a physical spot marketplace where buyers and sellers trade GPU capacity through reserve, forward, and refurbished hardware contracts. Pending CFTC approval, Silicon Data is partnering with CME Group to launch financially settled GPU futures and options, allowing participants to trade compute through existing futures brokers and clearing workflows. While CME futures are structured to settle financially, Compute Exchange also accommodates physical settlement via executable API calls for specific GPU clusters, such as verified B200 nodes in US East data centers. Li also piloted February- and April-settled contracts on Polymarket to assess speculative and retail interest, noting that market participants will ultimately adopt the trading venues and financial structures that best match their operational requirements. Standardizing Heterogeneity and the 'GPU Lottery' (9:07) A primary challenge in commoditizing compute is the lack of physical fungibility across hardware. Citing joint research with Jefferson Lab examining NVIDIA A100 chips, Li reports that identical hardware models exhibit up to a 38% performance variance due to silicon manufacturing differences, system architectures, and provider configurations. To counter this 'GPU lottery,' Compute Exchange acts as an independent verifying body that benchmarks nodes on FLOPS, memory bandwidth, token generation speeds, and service-level agreements before contracts are delivered, allowing buyers to balance verified performance against price and location. To construct reliable derivatives indices, Silicon Data normalizes diverse system setups rather than relying on simple price averages. Ingesting over 150,000 traded prices daily across more than 100 sources, the platform isolates auxiliary hardware variations such as CPU pairings, RAM, and storage specifications. This normalization produces a daily volatility between 20% and 30% for A100 and H100 indices—comparable to established commodities—contrasting with the 8% to over 100% volatility observed in unnormalized individual instances. Commercial Hedging and Market Participant Alignment (20:00) The compute market reflects structural dynamics seen in energy trading, dividing participants into naturally long and short exposures. Hardware owners, data centers, and NeoClouds with GPUs on their balance sheets are naturally long compute, requiring short futures positions to lock in stable revenue. Conversely, AI developers and enterprises requiring capacity for training and inference are naturally short compute; relying on unhedged on-demand markets leaves them exposed to price surges that historically swing from $3 to $6 or $9 per GPU hour. Consequently, market participants are shifting toward reserve and forward contracts to fix future costs. Li explains that establishing a liquid market requires financial speculators to take the opposing side of commercial hedging flows. However, she notes that compute trading demands domain expertise, as compute markets do not behave like power markets: software-layer advancements in model optimization and compression can suddenly enhance the utility of older chip generations, altering supply and demand fundamentals. Hardware Resale Trajectories and Macro Supply Constraints (28:40) Empirical transaction data contradicts narratives of rapid chip obsolescence and immediate market oversupply. Resale figures show refurbished NVIDIA H100 GPUs retaining roughly 85% of their value in year two and 84% in year three, while older enterprise L40S chips maintain rental yields around $0.40 per hour on hyperscale platforms. Rental rates for H100s stabilized and increased 10% to 15% over three months to approximately $8 per hour, while B200 rates rebounded above launch levels. Concurrently, Silicon Data's weighted LLM token index doubled to $2.21 per million tokens across 300 tracked models. Evaluating AI bubble concerns, Li analyzes machine-level cash flows rather than corporate equity valuations. Hardware investments remain commercially sound when discounted cash flows from binding multi-year forward contracts cover upfront purchase costs. Furthermore, headline-grabbing data center announcements face severe physical deployment bottlenecks—including power interconnect queues, co-location space, and optical fiber availability—which prevent sudden supply floods from destabilizing spot market pricing. While Compute Exchange and Silicon Data have developed the benchmarking and verification infrastructure for compute trading, the derivatives market remains dependent on CFTC regulatory approval. Long-term market depth will depend on whether financial speculators enter in sufficient volume to absorb commercial hedging demands as hardware generations and software efficiencies continue to evolve.

Carmen Li阐述了将其创办的Silicon Data与Compute Exchange推向市场的战略构想，主张通过建立标准化的定价指数与芝加哥商品交易所（CME）算力期货，将高度异构的GPU算力转化为类似原油的可对冲大宗商品。 算力金融化与供需两端的对冲机制 (3:52) 由Don Wilson支持的Silicon Data与Compute Exchange分别承担算力金融化的指数构建与现货履约职能。Silicon Data于2024年4月成立，并在2025年通过彭博终端发布了全球首个GPU定价指数，目前正推进待美国商品期货交易委员会（CFTC）批准的CME官方GPU期货与期权合约；Compute Exchange则作为现货市场，通过储备合约、远期合约及翻新硬件合约撮合交易。根据Li的观察，当前算力买方包括AI初创公司、传统企业及推理提供商，纯即时按需采购面临每小时单价从3美元骤升至6至9美元的剧烈波动，推动市场转向远期锁定。 Li将算力市场参与者与传统石油行业进行类比：持有GPU资产的NeoClouds及数据中心在商业结构上“天然做多算力”（类似Shell等石油巨头），需要通过做空期货合约锁定未来的稳定营收；而模型开发商与企业买家则“天然做空算力”，需要买入期货或看涨期权以规避硬件租赁成本的非预期暴涨。这种结构性风险敞口为算力衍生品市场奠定了内生交易需求。 解决“GPU彩票”与算力非同质化瓶颈 (9:07) 阻碍算力大宗商品化的核心障碍是其物理非同质性。Li援引其与Jefferson Lab联合发表的研究指出，由于芯片硅料制造差异、云服务商内部配置以及跨服务商系统架构的不同，即便在完全相同的NVIDIA A100（40GB显存）芯片之间，也存在高达38%的性能差异，这种现象被称为“GPU彩票”。为确立交割标准，Compute Exchange扮演了类似大宗商品普氏能源（PLATTS）或二手车评估系统CARFAX的角色，在合约交付前对算力节点的FLOPS、显存带宽、Token生成速度及SLA服务指标进行独立基准测试，允许买方按价格、地理位置或延迟溢价进行定制化采购。 在指数构建层面，Silicon Data通过吸收超过100个数据源、日均逾15万条实际成交价格（累计追踪约200家供应商的800多万个数据点），将涵盖不同CPU、内存和存储配置的异构数据标准化为统一基准，进而计算每日结算价。经标准化处理后，A100与H100指数的日波动率维持在20%至30%之间，与成熟大宗商品市场相当，有效压缩了非标算力可能产生的极端价格偏离。 CME结算架构与专业衍生品流动性 (22:45) 在交易与结算路径上，Li确认与CME合作的算力期货将遵循传统原油期货的先例，采用现金金融结算机制。市场参与者无需搭建新型金融基础设施，即可通过既有的期货经纪商与主经纪商账户，在现有的CME清算和保证金优化框架下开展业务；若未来在Compute Exchange引入物理交割，其形式也将表现为调用特定数据中心特定级别服务器集群（如美东机房A级B200）的可执行API访问权限。此前，Silicon Data还曾与Polymarket合作试行过在2月和4月结算的算力预测合约。 针对市场交易生态，Li强调成熟的期货市场不仅需要实体企业对冲，更需要金融投机者提供双向流动性以平衡风险。然而，算力交易要求极强的技术领域认知：芯片代际之间的相关性（如A100与H100）并非静态不变，且软件层面的算法压缩与模型优化进展能够动态激活旧代芯片的可用性，从而在根本上重塑硬件供需关系，传统大宗商品交易员无法直接机械套用旧有逻辑。 二手芯片残值与机器级现金流评估 (28:40) 针对算力硬件快速贬值的市场担忧，Li给出的实际交易数据呈现出更平缓的折旧曲线。统计表明，翻新H100芯片在第二年的转售残值约为原始价格的85%，第三年仍能维持在84%左右，贬值速度显著慢于传统乘用车；即使是上一代L40S芯片，在超大规模云平台上仍能保持每GPU小时约0.40美元的稳定回报。此外，Silicon Data追踪约300种大语言模型的Token综合指数近期达每百万Token 2.21美元；在硬件租赁端，B200在经历短暂回撤后反弹突破首发价，而H100租赁单价在三个月内上涨10%至15%至约8美元/小时。 在评估所谓“AI基础设施泡沫”时，Li主张应脱离风险投资估值叙事，回归机器级别的现金流贴现模型。硬件运营商若能通过签署多年的远期算力合约覆盖采购CapEx支出，资产在财务层面即具备可持续性。同时，即便有公开宣布的250亿美元级超级数据中心项目，实际算力上线仍严格受制于机房空间、电力互联和光纤网络等物理基础设施瓶颈，现实工程约束有效防范了瞬时供给过剩对价格的冲击。 Carmen Li认为通过基准指数与CME金融衍生品能够实现算力资源的平稳对冲与定价透明，但该体系的最终效能仍依赖于监管审批进展、投机者跨领域定价能力以及模型优化对底层硬件寿命曲线的持续扰动。

## Background 本期节目由彭博社（Bloomberg）旗下的《Odd Lots》播客于 2026 年 6 月 15 日发布，录制于纽约 City Winery 的线下专场。节目简介指出，继此前专访探讨“GPU 市场体量或将超越原油”后，主持人邀请到了与 DRW 创始人 Don Wilson 携手创业的 Carmen Li（同时掌舵 GPU 指数编制机构 Silicon Data 与现货撮合平台 Compute Exchange），深入探讨如何建立算力指数、应对算力非同质化挑战，以及构建标准化的 GPU 期货市场。 ## TL;DR Carmen Li 深入剖析了将 GPU 算力金融化为可交易大宗商品的完整构想，通过 Silicon Data 编制彭博终端价格指数并与 CME 合作研发算力期货，同时依托 Compute Exchange 撮合现货与远期算力采购。针对同款芯片性能差异高达 38% 的“GPU 彩票”非同质化痛点，平台引入类似 CARFAX 的第三方跑分验证与多因子回归模型，将非标算力归一化为基准配置以实现精准对冲。面对所谓“芯片两年即彻底报废”的舆论恐慌，她用真实数据证明 H100 在第二、三年仍保留约 85% 的高残值，并指出短期内电力、托管机房与光纤等物理交付瓶颈让算力无法无序过剩，标准化的 CME 期货正是重资产企业对冲价格波动风险的迫切解药。 ## Atmosphere & style 本场现场录制洋溢着华尔街交易员思维与前沿硅谷算力碰撞的火花，观众与台上的互动兼具临场感与幽默感。主持人 Tracy 与 Joe 频频从经典大宗商品（如原油对冲、基差交易与 CDS）的角度展开层层追问，而 Carmen Li 则以深入浅出的汽车行业比喻（如 CARFAX、二手车折旧）、扎实的实测图表和风趣坦诚的表达见招拆招。整场对话既有勾勒未来算力衍生品帝国的宏大金融叙事，又充满了对机架、电力配额和光纤等物理落地难题的务实敬畏，智识密度极高且节奏明快。 ## Bulletin - `05:00` Silicon Data 已在彭博终端上线首套 GPU 价格指数，并与芝加哥商业交易所（CME）合作，计划在获得 CFTC 批准后推出以现金结算的 GPU 期货与期权产品。 - `10:00` 联合杰斐逊实验室的测试显示，相同硬件标称的英伟达 A100（40GB）在不同环境下的实际性能差异高达 38%，使得算力交付面临“GPU 彩票”现象，亟需独立第三方基准测试进行同质化校准。 - `25:30` 真实消耗量加权的百万 LLM Token 混合成本达到约 2.21 美元，较去年 12 月翻倍；同时英伟达 B200 租金在下调后罕见暴涨突破上市初期高点，老款 H100 与 A100 也逆势走高。 - `30:00` 驳斥了芯片使用寿命极短的悲观论调，实测数据显示翻新英伟达 H100 第 2 年残值保持在原价的 85%，第 3 年保持在 84%，展现出远高于传统汽车与 IT 硬件的保值率。 ## Key takeaways - `00:00` 算力金融化正在形成双层结构：Compute Exchange 负责撮合现货、预留及翻新等物理算力采购，而 Silicon Data 则提供基准指数并对接 CME 开发衍生品。 - `06:45` 算力市场高度映射传统原油生态：拥有重资产服务器的 NeoCloud 属于“自然多头”，需防范租金下跌；而软件及 AI 初创等算力消耗方是“自然空头”，迫切需要锁定用算成本。 - `08:30` “GPU 彩票”现象揭示了芯片的非同质化困境，即使同代同款芯片，其性能也会因机架拓扑、基础设施和固件调度产生巨大差异。 - `10:00` Compute Exchange 通过打造类似“GPU 版 CARFAX”的机制，在交付前提供涵盖 FLOPS、显存带宽和 Token 吞吐量的独立验证报告，消除信息不对称并支持差异化定价。 - `11:15` 构建算力指数不能直接进行简单算术平均，需借助多因子回归模型消除 CPU、内存、互联带宽与地理位置带来的溢价，将非标数据归一化为基准配置模型。 - `13:30` 主流芯片经过标准化处理后的日度波动率维持在 20% 至 30% 之间，属于成熟大宗商品极具流动性与对冲价值的健康区间，而非外界所误解的单边永远涨或永远跌。 - `16:40` Silicon Data 穿透传统云服务商的私密双边市场（OTC），通过 API 和 S3 直传等方式整合了全球约 200 个数据源及 800 万个活跃价格点。 - `22:38` CME 算力期货将采用现金交割机制，让传统大宗商品交易员和机构可以直接沿用现有经纪账户与保证金体系无缝入场。 - `22:38` 在物理现货层面，未来的算力交割将以 SiliconMark 基准测试为依归，最终通过规范化的 API 调用凭证实现交付，而非寄送实体服务器晶圆。 - `25:30` 当前模型推理算力市场需求极度旺盛，覆盖 300 多款开源及闭源模型的数据表明，混合 Token 成本不仅没有随技术成熟下降，反而自去年底以来翻倍。 - `26:50` 最新一代 B200 租金在产能释放阶段打破历史规律逆势暴涨，且老款硬件租金被全线托底，表明算力供给紧缺正沿产品梯队自上而下蔓延。 - `29:15` 二手芯片翻新正在兴起“Carvana 模式”，帮助拥有机房与电力的 NeoCloud 规避长交付周期，通过快速部署二手 H100 抢跑变现以最大化资本回报率。 - `30:00` 从贴现现金流角度看，翻新 H100 运行到第三年依然能维持 84% 的原价残值，彻底推翻了“AI 芯片两年即报废清零”的过度悲观叙事。 - `32:20` 衡量 AI 是否存在泡沫应立足于底层机器的未来折现现金流能否覆盖购机成本，而短期内“大规模算力过剩”是伪命题，因为电力、光纤和机位配额等物理交付瓶颈极高。 - `35:00` 场外长期租赁协议的违约风险属于标准信用风险，而当前整个产业链最缺乏的是针对底层算力价格波动的标准化对冲机制，这也正是 CME 期货的核心价值所在。 ## Detailed walkthrough ### `00:00` 播客开场与“算力能否成为下一个原油”的探讨 本期节目依然来自在纽约 City Winery 举办的现场录制专场。主持人 Tracy Alloway 和 Joe Weisenthal 谈到了节目的延续性——早在约六七个月前的芝加哥现场秀中，他们就曾与 DRW 创始人 Don Wilson 展开过深度对话，当时 Wilson 重点谈到了高频交易环境以及他正在押注的一个全新前沿领域：构建一个体量可能超越原油市场的 GPU 算力市场。在当前的 AI 浪潮下，GPU 算力作为极端稀缺的关键资源，引发了整个金融界与科技界的强烈关注。 大家最核心的疑问在于：稀缺的 GPU 算力能否真正演变成一种像原油一样具备高度流动性的可交易大宗商品？企业是否能通过购买期货合约来锁定未来使用算力的成本？买家能否转售这些期货合约？市场上又是否会出现专门针对某款主流芯片（比如英伟达 H100 运行一小时的价格波动）进行多空投机的专业投机者？虽然业界普遍热衷于使用“算力是新时代的原油”这一比喻，但建立这样一个新型市场面临着严峻的结构性挑战，其中最大的障碍就是“同质化”（Fungibility）问题——同一款芯片由于网络拓扑、固件版本或所在数据中心的不同，其实际性能输出可能天差地别，彼此之间根本无法简单等同。为了解开这些构建设想背后的谜团，本期节目邀请到了 Don Wilson 的创业合伙人、同时掌舵两家核心公司的女性 CEO Carmen Li。 ```mermaid flowchart TD accTitle: 算力金融化市场结构 accDescr: 展示实体算力采购与衍生品市场之间的双层映射结构 subgraph PhysicalLayer ["实体算力层：Compute Exchange"] A["算力提供方<br/>(NeoCloud / 算力持有者)"] -->|"现货 / 预留 / 远期合约"| B["算力需求方<br/>(AI初创公司 / 传统企业 / 推理服务商)"] end subgraph FinancialLayer ["金融衍生品层：Silicon Data & CME"] C["底层数据与指数构建<br/>(Silicon Data)"] -->|"价格基准数据"| D["算力期货与期权<br/>(CME 衍生品合约)"] D -.->|"对冲成本与收入波动"| A D -.->|"锁定未来算力价格"| B end ``` ### `05:00` 算力双子星：Silicon Data 与 Compute Exchange 的定位 Carmen Li 回忆起半年前主持人采访 Don Wilson 时的情景，当时正值八九月份，市场上各类芯片的算力租赁价格全线走低，而如今价格再度飙升，正印证了此前对市场周期与算力价格波动的讨论。Carmen Li 详细梳理了自己统领的两家关键公司： 1. **Silicon Data**：担任算力资产的指数编制机构（Index Provider）。该团队自 2024 年 4 月起便潜心研发 GPU 价格指数，并在 2025 年于彭博终端（Bloomberg Terminal）上线了全球首套 GPU 价格指数。目前，公司已与芝商所（CME）正式达成合作，正在等待美国商品期货交易委员会（CFTC）的监管批准，计划在数月内正式推出基于该指数的 GPU 期货与期权产品。 2. **Compute Exchange**：实体算力采购的现货撮合平台（Spot Marketplace）。主要面向市场提供预留合约（Reserve Contracts）、远期交付合约（Forward Contracts）以及翻新/二手硬件合约（Refurbished Contracts），解决当下实打实的物理算力交割与采购需求。 ### `06:45` 现货买家画像与原油对冲模型的映射 针对主持人关于“谁会去现货市场上购买算力”的提问，Carmen Li 剖析了算力市场参与者的丰富构成与结构演变。在供给端，平台聚集了遍布全球的新型云服务商（NeoClouds）；在需求端，结构则极其多元： - **AI 初创企业**：虽然公司规模尚小，但在模型训练与研发上动辄每月烧掉数百万美元的 GPU 费用。 - **传统企业（Enterprises）**：它们并非底层大模型开发商，但在业务中需要部署一到两个计算节点、数台服务器来进行推理（Inference）或特定业务管线集成。 - **无算力资产的推理服务商（Inferencing Providers）**：自身不拥有重资产服务器，但需要采购算力以为其他终端应用托管开源、开放权重的模型服务。 在合约形式上，虽然按需付费（On-demand）看似提供了“不用即不付钱”的最大灵活性，但买家却会完全受制于供需曲线的极端震荡——芯片每小时的价格可能会瞬间从 3 美元跳涨至 6 美元甚至 9 美元，不仅让企业的业务利润率无法预测，而且在算力告急时根本无法保证下个小时或下个月还能抢到资源。因此，越来越多的机构转向提前锁定未来特定月份交付的远期合约（如锁定 9 月、11 月或来年 2 月的资源）。 Carmen Li 将这一生态巧妙地类比为经典的原油市场（WTI 与 Brent）： - **自然多头（Naturally Long）**：在原油市场是类似壳牌（Shell）这样坐拥油井的开采商；在算力市场则是重资产拥有服务器集群的 NeoCloud，或是将 GPU 列入资产负债表作为质押资产的金融机构。它们为了防止算力暴跌侵蚀收入，天然具有在期货市场做空或买入看跌期权以锁定收入的对冲需求。 - **自然空头（Naturally Short）**：在原油市场是像美航（American Airlines）这样极其依赖航油的企业；在算力市场则是几乎所有需要消耗算力的软件公司和终端用户。它们需要借助期货合约来控制成本波动率，锁定未来算力开支。 ### `08:30` 算力品质的非同质化陷阱与“GPU彩票” 当主持人 Tracy Alloway 追问在现货市场交易时，买家如何确保拿到的算力符合质量预期、算力是否真能做到完全无差异交付时，Carmen Li 打开了准备好的演讲幻灯片，抛出了一个核心概念——**“GPU 彩票”（GPU Lottery）**。她指出，不同于桶装原油拥有相对明确的 API 度数和含硫量等理化指标，物理 GPU 算力的非同质化问题远比想象中复杂。团队早在年初的行业会议（如 GTC 等前沿计算峰会）上就专门发表了关于该现象的研究论文，揭示即便是同一批次采购、同一数据中心上架的同规格芯片，在实际性能表现和稳定性上也存在显著差异，使买家在缺乏标准化基准时如同在“抽彩票”。 ### `10:00` 算力性能差异与“GPU 版 CARFAX”的独立验证机制 在讨论算力如何成为标准化大宗商品时，卡门·李（Carmen Li）抛出了一个令人震惊的研究数据：她与杰斐逊实验室（Jefferson Lab）联合进行的一项测试显示，哪怕是同一款芯片——具体测试对象为拥有 40GB 显存带宽的英伟达 A100——其性能差异居然高达 38%。研究团队将这种差异逐层拆解为芯片硬件本身固有的公差、同云服务商内部的系统架构差异（intra-provider），以及跨服务商网络与基础设施调度导致的外部差异（inter-provider）。卡门打趣道，台下观众在租用算力时根本无法预知自己拿到的是“神仙卡”还是“缩水卡”，正如二手车买家不看报告就提车一样充满未知的撞大运风险。 为了彻底消除买卖双方的信息黑洞，她的算力现货交易平台 Compute Exchange 试图扮演“算力界的普氏（Platts）”或“GPU 界的 CARFAX”。在该平台上，算力交付并非只是一纸模糊的合同，而是在实际交付给客户之前必须通过独立第三方的基准性能检验。如果客户发起采购意向（RFP），比如需要 200 个 B200 算力节点，平台不仅会给出硬件规格表和商业商务条款，还会同步出具一份独立验证报告，实测其真实的浮点运算能力（FLOPS）、显存带宽、每秒生成 Token 速率以及服务等级协议（SLA）等硬性指标。卡门强调，透明度和选择权才是核心：“作为最终用户，你可以明确决定价格是否是唯一的首要考量；或者，你是否心甘情愿为了更好的地理节点低延迟、更强的数据吞吐吞吐性能而支付溢价。” - **芯片性能方差实测**：即使在硬件标称完全一致的前提下，同一款 A100（40GB）在实际负载下的算力输出可能相差多达 38%。 - **全方位独立质检**：交付前必须实测底层硬件指标（FLOPS、显存带宽）、LLM 关键吞吐（Tokens/s）以及网络与 SLA 达标情况。 - **差异化定价策略**：让买家自由权衡价格优先还是质量溢价，打破传统算力市场“开盲盒”式的采购弊端。 ```mermaid flowchart TD accTitle: Compute Exchange 算力验证与定价流 accDescr: 展示从采购需求提出、多维度独立跑分验证到买家基于基准指标决策的完整流转过程。 subgraph Demands ["1. 买家发起意向"] A["买方发出需求 (如 200 个 B200 节点)"] end subgraph Verification ["2. 交付前独立跑分验证 (CARFAX)"] B["硬件算力测试 (FLOPS & 显存带宽)"] C["AI 负载测试 (Token 吞吐量 & 延迟)"] D["基建审计 (网络拓扑 & SLA 保障)"] end subgraph Decision ["3. 撮合与透明定价"] E["算力规格与价格基准透明化"] F{"买方根据溢价自主决策 (低价优先 vs 性能/地理优先)"} end A --> B & C & D B & C & D --> E E --> F ``` --- ### `11:15` 石油对标、基差风险与无法做简单平均的指数模型 当主持人提到原油市场存在布伦特（Brent）和西德克萨斯轻质原油（WTI）等全球公认的定价基准，并顺势提出疑问：当下各种 AI 专用芯片层出不穷（例如针对大模型推理优化的 Cerebras 晶圆级芯片、谷歌 TPU 等），一个通用的“H100 价格指数”是否足以作为整个异构算力市场的套期保值工具？卡门直言不讳地指出，这正是她创建指数公司 Silicon Data 的根本原因：“金融产品存在的唯一功能性意义，尤其是在大宗商品领域，就是为了套期保值（hedging）。投机当然也很好，但核心价值是让企业管理波动性、进行风险转移与资本配置；如果我们做出来的指数连对冲都做不到，那我们的工作就是彻底失败的。” 卡门解释称，构建 GPU 价格指数绝非把市场上两台 H100 节点的出租单价拉出来取个简单算术平均那么粗暴。因为每家服务商搭配的 CPU 架构、内存配置、本地固态硬盘读写速度、互联带宽乃至机房地理位置完全不同，直接平均无异于“拿苹果和橙子做对比”。Silicon Data 的指数建模会深入追溯超过 100 个数据源过去半年的真实交易历史，运用严苛的多因子回归模型甄别驱动价格溢价的真正推手。平台每天接入超过 15 万条真实成交价格，先将这些带有繁杂特征的非标交易统一折算归一化为“基准配置模型（Base Case）”，随后测算出统一的结算价。尽管这与用户在各大新兴云服务商（Neoclouds）账单上的实际支出高度相关，但卡门提醒，算力市场就像原油基差交易一样天然存在“基差风险（Basis Risk）”，比如部署在美东机房的算力成本可能天然偏高，而指数的作用正是帮量化交易员和算力买家精准测量和对冲这种基差。 - **金融对冲的核心职能**：指数工具的根本目标是帮助重资产投入者转移算力价格波动风险，而非仅供投机炒作。 - **拒绝简单算术平均**：单卡租金受制于 CPU 瓶颈、内存大小、存储吞吐与骨干网延时，非标特性决定了必须采用因子模型进行归一化。 - **海量数据流与基差核算**：每日摄取超 15 万笔成交价进行模型拟合与基线化折算，提供可量化管理的基差风险敞口。 --- ### `13:30` 算力价格的健康波动率：“总在跌”与“总在涨”的悖论 面对主持人关于“算力市场是否具备足够价格波动，以支撑起一个庞大衍生品市场”的质问，卡门忍不住分享了一个颇具讽刺意味的行业观察：去年行业算力供给出现暂时回暖、GPU 租赁价格连续下挫时，行业内外所有人都在质疑：“如果算力注定是一个价格永远下跌的硬件消耗品，建立价格指数有什么用？”然而戏剧性的是，到了今年大模型爆发、算力极度供不应求使得租金飙升时，蜂拥而来的客户又抛出相反的疑问：“如果价格注定永远上涨，我们干嘛还需要指数？”卡门对此哭笑不得：“这真是一件极其迷人的事情，这就是我每天要回答的所有问题。” 卡门进一步明确，金融市场衡量波动的标准从来不是单边行情的绝对涨跌幅，而是每日波动率的离散程度（Daily Volatility）。经过标准化清洗后，A100 和 H100 的日度波动率通常平稳维持在 20% 到 30% 之间，卡门评价这在成熟大宗商品市场中属于“非常健康的波动区间”。如果不进行科学的归一化处理，单一配置芯片在各个局部机房的价格波动甚至可能从极度死水微澜的 8% 骤变至狂暴的 100% 以上。正是指数对网络带宽、机房选址及配套硬件的标准化校准，才熨平了局部非理性杂音，将真实的算力供需波动沉淀为一个具有高度流动性与可交易性的成熟资产类别。 - **市场认知误区**：无论是单边单调下跌还是单调上涨，业界往往忽略了宏观技术迭代与微观周期波动交织带来的对冲需求。 - **大宗商品级波动表现**：主流芯片经过归一化后的日度波动率落在 20%~30%，完美契合衍生品交易与现货套保的流动性要求。 - **微观极端波动的平抑**：未经标准化的单点芯片价格波动极度扭曲（8% 至 100%+ 不等），指数化提取出了纯粹的市场供需信号。 --- ### `16:40` 穿透云厂商黑盒：从零搭建 800 万级定价数据管道 当被问及彭博终端屏幕上跳动的数据究竟如何“从无到有”时，卡门谈到了算力大宗商品化道路上最艰辛也最接地气的脏活累活——算力数据采集。与传统场内撮合交易的大宗商品不同，现存的 GPU 算力现货交易充斥着大量双边私下成交（OTC）。除 Compute Exchange 自身的撮合数据外，各大超大规模云服务商（Hyperscalers）和新兴算力云（Neoclouds）在合约期限、机型定制化程度、保密协议（NDA）及接口开放度上千差万别。曾在彭博数据部门深耕多年的卡门深知底层数据源的多元化才是指数可信度的生命线，因此她的工作很大一部分就是逐家与这些云巨头打交道，展开一场场冗长的商业谈判与数据授权拉锯。 “你首先得找上门去，跟对方聊：‘嘿，我非常欣赏你们的业务，你们是极棒的新型云，我能获得你们平台交易数据的授权吗？’对方的第一反应往往是：‘这对我有什么好处？’紧接着他们会担忧：‘如果我把全部底牌数据都交给你，商业机密不全泄露光了吗？’”卡门幽默地描述了这种谈判日常。为了消除顾虑，团队必须起草极其完备严格的脱敏授权协议，逐一敲定管线接入细节：究竟是通过 AWS S3 存储桶直传脚本定期推送，还是双方互相对接 API 接口。如今，这套原本被外界视为天方夜谭的私有数据获取机制已变成团队的日常运营例行公事（BAU）。Silicon Data 已经连接了全球约 200 个数据源，并拥有高达 800 万个活跃价格点。卡门笑着感叹，行业里现在只要提到她，大家的第一印象就是：“卡门那家伙准又来找我要数据了。” - **穿透私密双边市场**：通过标准化法律架构与商业利益绑定，突破新兴云厂商因担忧商业机密泄露而产生的信息壁垒。 - **工程级数据管线部署**：支持从定制 S3 Bucket 批量作业到低延迟双向 API 调用的多样化安全摄取方案。 - **全球数据规模突破**：成功集成约 200 个跨云服务商数据源，汇集逾 800 万个全球定价数据点，构筑起底层指数的数据护城河。 ### `20:00` 数据方法论的分歧与算力期货交易者的入场 对于市场上不同算力价格指数之间的差异，Carmen Li 表示自己无法评价竞争对手的方法论，但底层数据与计算方法的不同必然会导致最终价格分歧。她强调，在金融市场中“市场永远是用脚投票的（the market always votes）”。一旦衍生品开始交易，流动性自然会涌向最能够真实反映风险、最不易被操纵、最能帮助实体企业进行套期保值的基准。 在讨论市场参与者结构时，主持人指出，任何成熟的衍生品市场不仅需要持有算力空头敞口的套期保值者（如急需算力的 AI 企业），同样需要承担风险的投机者与做市商。Carmen 透露，各大投行、市场做市商和独立投机机构对此表现出极高的兴趣，并已沟通了相当长时间。部分大型银行内部拥有双边资金流，可以通过内部头寸相互冲销，而更多机构则需要外部杠杆产品来构建敞口。从交易所的角度，Carmen 幽默地坦言：“出于私心，交易的人越多对我自然越好。” 但她同时也提醒，算力交易绝非把传统交易员直接从原油或电力交易台平移过来那么简单。理解 GPU 交易需要极高维度的专业认知： - **异质性资产**：GPU 并非同质化大宗商品（homogeneous product），不同架构与代际的芯片不能简单划等号。 - **关联度动态演变**：当前 A100 与 H100 的相关性并不高，某些应用场景彼此割裂，但未来是否可互相替代仍存变数。 - **软件层的乘数效应**：随着模型量化、模型压缩技术与推理优化的演进，今天只能部署在最新旗舰芯片上的大模型，六个月后或许就能跑在老一代芯片上，这正在剧烈重构局部的供需曲线和终端用户行为。 --- ### `22:38` 结算机制与交易接入：从现金交割到 API 交付 针对算力期货如何交割的疑问，主持人打趣是否会有人像交割原油一样，收到装满整张服务器晶圆的实物包裹。Carmen 笑着回应，与芝加哥商业交易所（CME）合作推出的算力期货将采用与传统原油期货相同的**现金结算（financially settled）**方式。不过在 Compute Exchange 现货市场端，远期合约与实物履约一直保持开放。借助其开发的基准测试工具 SiliconMark，未来的物理交付场景甚至可以精确规定等级、配置与地理区位（例如指定位于美东、符合特定性能规范的 B200 服务器集群），而交割的最终形式通常是 API 访问调用凭证，而非真的寄送硬件。 ```mermaid flowchart TD accTitle: 算力衍生品与现货结算流程架构 accDescr: 展示了 CME 金融结算与 Compute Exchange 物理/API 交付的双轨交易架构。 subgraph CME["CME 算力期货市场"] A["标准期货合约"] --> B["传统大宗商品经纪账户"] B --> C["现金交割 (Financial Settlement)"] end subgraph CX["Compute Exchange 现货与远期"] D["SiliconMark 算力基准测试"] --> E["标准化服务器规格 (如 US-East B200)"] E --> F["API 算力访问调用交付"] end Market["交易参与者"] --> CME Market --> CX ``` 至于交易的准入门槛与操作路径，Carmen 解释称，依托 CME 的成熟金融基础设施，交易算力期货与交易现存的 CME 产品没有任何区别。投资者无需下载新软件或开立陌生账户，原有的经纪商平台、清算渠道以及保证金冲抵机制完全通用（BAU - Business As Usual），这极大降低了机构的入场摩擦。 当被问及预测市场是否构成分流竞争时，Carmen 透露去年有人在未经许可的情况下将她的产品列上了 Polymarket。他们随后主动与 Polymarket 团队对接，在 2026 年 2 月和 4 月试点上线了几个交割周期的实验性合约来测试水温。尽管目前在衍生品方面与 CME 保持独家合作，但她对预测市场的探索持完全开放的态度，坚信只要合规授权，最终市场会挑选出体验最优秀的产品。 --- ### `25:30` 推理市场的真实全貌与代币定价指数 谈及生成式 AI 的算力消耗格局，市场目光往往局限于 OpenAI、Anthropic 和 Google 这几家头部闭源巨头。然而 Carmen 强调，真实的推理（Inference）算力世界远比公众认知广泛。Silicon Data 持续在全球范围内追踪超过 300 款开源、开放权重（open weights）以及闭源模型的定价与消耗态势。 虽然尚未正式公开发布大语言模型代币（LLM Token）价格指数，但彭博终端上已可查阅相关数据。指数显示，目前基于真实消耗量加权的一百万 Token 混合成本（包含输入与输出 Token）约为 2.21 美元，相较于去年 12 月 1 日的水平已经**翻倍**。这种结构性反弹印证了模型推理端算力需求的爆发式扩张，且不再单纯受少数几家模型实验室的主导。 --- ### `26:50` 现场图表解读：新老 GPU 租金的反常暴涨 在现场演示的大屏幕上，Carmen 拆解了一张核心走势图，该图记录了三款英伟达芯片在新型云服务商（NeoCloud）中每 GPU 每小时的按需租赁价格变动： | 芯片型号 | 历史常态价格规律 | 近期实测走势异动 | 供需信号解读 | |:--- |:--- |:--- |:--- | | **B200**（黄线） | 新品发布初开盘价高，随后随产能释放快速走低并平稳 | 初期价格如期回落，但下降斜率比以往更平缓，随后迅速掉头暴涨，突破上市初期的历史高点 | 需求远超产能扩张速度，处于与以往周期截然不同的超紧平衡阶段 | | **H100**（红线） | 经过产能释放后价格走低并进入平台期 | 去年经历适度降价后强劲反弹，过去三个月内小时租金上涨约 10% 至 15%，回升至约 8 美元/小时区间 | 尽管不再是顶级旗舰，但作为主力计算卡，受到推理算力缺口的有力托底 | | **A100**（更早一代） | 沦为低端大宗商品（Commodity），长期承压 | 价格曲线触底反弹，同样出现温和上扬趋势 | 市场向后兼容、追求性价比方案的算力外溢效应明显 | Carmen 重点分析了 B200 的反常表现：传统上，每一代新硬件（如早期的 A100、H100）问世时价格都处于高位，随后伴随供应链产能就位，租金斜率通常呈现较为陡峭的下坠，并在某一稳态区间沉淀。然而 B200 在今年早些时候下调后，下跌幅度极其有限，旋即以惊人的速度向上攀升，价格甚至超越了最初发布时的峰值。而作为老款硬件的 H100 与 A100 价格逆势回涨，更说明算力短缺正在沿产品梯队自上而下漫延。 --- ### `29:15` 二手算力翻新：数据中心领域的“Carvana”模式 随着算力缺口扩大，Compute Exchange 也延伸出了二手芯片翻新（Refurbishment）业务，被主持人形容为算力市场的“Carvana 汽车置换模式”。 Carmen 解析了这项业务背后的商业逻辑：许多新型云服务提供商（NeoCloud）好不容易锁定了土地、电力配额以及主机托管数据中心（Co-location），这时他们将面临一道典型的资本回报率（ROI）抉择题： - **选择等待**：向原厂预订最新一代的 B300 或 GB 系列整机柜，但必须承担长达数月乃至更久的交付排队周期，期间电力和场地租金持续沉没； - **翻新即用**：采购经过官方检测翻新的旧代际 GPU（如 H100 甚至更早芯片），立即上架通电点亮，在当前租金回暖的市场环境中快速抢跑变现。 对于这些数据中心运营商而言，这完全是一场关于时间价值与投资回报周期的量化精算。翻新芯片的低门槛与快速交付能力，为算力基础设施的资本流转提供了第二种解题思路。 ### `30:00` GPU 残值评估与芯片寿命的迷思：85 折与折旧真相 在评估 GPU 的残值（residual value）以及二手翻新交易时，Carmen Li 指出，这本质上是一个非常纯粹、基于现金流的投资回报率（ROI）测算。以英伟达 H100 为例，尽管它的出租费率显然无法与更新型的 B200 媲美，但其硬件采购成本基数也更低。如果企业锁定一份为期三年的远期出租合约，再将未来的现金流折现至今，就能明确得出这台硬件当下的残值。Silicon Data 正是通过这种计算帮助客户理解 GPU 能创造的真实价值，以及二手翻新市场上公允的交易价格。虽然翻新芯片需要经过严格的软硬件测试以确保其正常运作，存在不少技术细节，但底层逻辑完全是财务驱动的。 这一残值模型直接切中了去年年末市场上围绕“AI 芯片寿命”的狂热焦虑。主持人提到，大约在去年 11 月至 12 月，知名做空投资者迈克尔·伯里（Michael Burry）在社交媒体上发帖质疑芯片的使用寿命，随后整个金融市场陷入了长达三周的恐慌式讨论，之后又像什么都没发生一样迅速翻篇。Carmen 笑称，当时也有记者就此采访过她，虽然她自嘲“当时人微言轻，现在也一样不重要”，但她早在博客中用真实数据反驳了这种过度悲观的叙事。 根据 Carmen 当时跑出的残值曲线数据，硬件进入第二年时，翻新 H100 芯片的转售残值大约还能维持在原本价格的 85%（85 cents on a dollar）；到了第三年，残值依然坚挺在 84% 左右。Carmen 打趣道：“我开自己的车开了快十年，汽车的折旧贬值幅度可比这个夸张多了！”数据显示，GPU 价格在第一年到第二年之间确实会出现一定幅度的自然下滑，但随后曲线便趋于平缓。即使到了更早一代的显卡（如被戏称为老兵级的 L40S），云厂商巨头（Hyperscalers）如今依然按每小时每卡 0.40 美元的价格向租户收费。Carmen 直言：“我不知道市场上所谓‘两年就彻底报废贬值’的数据是从哪里编出来的，如果有人愿意在第二年以一折（10 cents on a dollar）的价格把二手芯片甩卖给我，我愿意每天都全盘接下这笔交易。” - **H100 残值保留率**：第 2 年二手翻新价格约为原价的 85%，第 3 年约为 84%，展现出远超传统 IT 设备和乘用车的价值稳定性。 - **上一代显卡商业价值**：老一代 L40S 至今仍在主流超大规模云厂商处稳定出租，费率维持在 0.40 美元/GPU/小时。 - **透明度工具**：Silicon Data 已将测算代码向数据订阅客户开放，客户可直接运行底层代码实时获取最新的残值模型测算。 --- ### `32:20` 机器层面的“AI 泡沫”定义与供应链物理落地的重重瓶颈 面对各界始终争论不休的终极问题——“我们是否正处于一场 AI 泡沫之中”，Carmen 保持了指数编制机构的严谨态度。她强调，作为指数提供商，公司不仅有合规免责声明的要求而不能提供前瞻性指引，更坦承“谁也无法百分之百预知未来”。不过，她从基本面资产估值的角度给出了清晰的框架：正如当年纳斯达克指数飙升 200% 后又暴跌 84% 构成了典型的资产泡沫一样，判断泡沫的核心在于“未来的现金流能否支撑当下的估值”。与风险投资（VC）评估 OpenAI 等初创企业高耸的股权估值不同，Carmen 聚焦的是物理世界中的 GPU 机器。 在硬件资产层面，测算泡沫其实相当直截了当：手握远期合约预期的未来现金流，通过折现计算，看硬件产出能否收回购机成本即可。针对市场普遍担心的“如果未来 AI 需求雪崩、或者行业过度建设（Overbuild）导致供过于求、租金价格暴跌怎么办”，Carmen 分析指出，尽管未来总需求的演变难以精准预测，但所谓的“大规模算力过剩”往往被舆论夸大了，因为让算力真正上线运行存在极高的现实物理壁垒。 媒体上动辄充斥着各大科技巨头宣布投资 250 亿美元自建超大规模数据中心的新闻，但这些资本支出承诺绝不等于现成的可用 GPU 算力。现实中，客户购买全新的服务器需要经历漫长的排队等待（waitlist），更要面对机房托管机位配额（co-location）、电力接入以及铺设高带宽光纤（optic fiber）等一系列严苛限制。算力从图纸走向机架的链条极其繁复，“所有的星象都必须完全对齐”（all stars have to align），因此短期内并不会凭空出现大量冲击市场的无序供应。 ```mermaid flowchart TD accTitle: GPU 算力落地交付流程链条 accDescr: 展示从资本投入到 GPU 算力最终上线所需跨越的多个硬性物理基础设施瓶颈。 subgraph Capital["资本与订单阶段"] A["资本支出宣布<br/>(如 250 亿美元数据中心规划)"] --> B["订购全新 GPU 服务器<br/>(面临漫长的供货排队名单)"] end subgraph Infra["物理基础设施交付瓶颈"] B --> C["机房托管空间 (Co-location)<br/>与电力配额锁定"] C --> D["高带宽光纤网络部署<br/>(Optic Fiber)"] end subgraph Delivery["实际可用算力"] D --> E["真实可用 GPU 算力并网上线<br/>(受多重物理约束，释放周期漫长)"] end ``` --- ### `35:00` 违约风险的边界：为何算力市场迫切需要 CME 期货对冲 针对 Carmen 提到的“长期租赁协议锁发现金流”，主持人追问了一个现实痛点：如果签署了长期合约的客户中途违约怎么办？算力市场未来是否会出现类似于信用违约互换（CDS, Credit Default Swaps）的衍生品来防范交易对手风险？ Carmen 回应道，交易对手违约的风险存在于任何场外交易（OTC）市场，无论对方是多么知名的大机构，在场外双边交易中都无法完全规避这种信用风险，金融界也已经拥有许多常规的风险控制与信用对冲工具来管理此类敞口。然而，当前算力买家与数据中心运营方真正**无法**对冲的，是入场锁定资产时的 GPU 价格风险与底层算力成本波动。 正因如此，通过芝加哥商业交易所（CME）建立标准化的 GPU 期货合约才显得至关重要。标准化的期货交易所能够提供场外市场无法比拟的价格透明度与充沛流动性，使算力生态中的硬件买家、数据中心运营商和租赁方能够以极低的交易摩擦便捷地建仓和平仓，从而真正对冲算力采购与资本投入的价格下行风险。在 Carmen 详尽解答后，主持人 Tracy Alloway 与 Joe Weisenthal 向其表达了由衷的感谢，本场录制于纽约 City Winery 的 live show 访谈也在热烈的掌声中顺利收官。 - **信用违约与价格风险的区隔**：场外租赁合约的违约风险属于标准信用风险，可沿用传统风控手段；但 GPU 本身的重置成本与出租价格波动，则需要衍生品工具对冲。 - **CME 期货的核心价值**：解决目前计算资源现货与场外合约中流动性碎片化的问题，赋予产业参与者像管理原油、天然气或大豆风险一样对冲算力敞口的能力。 ## Notable quotes > `10:00` 作为最终用户，你可以明确决定价格是否是唯一的首要考量；或者，你是否心甘情愿为了更好的地理节点低延迟、更强的数据吞吐性能而支付溢价。 — Carmen Li > `11:15` 金融产品存在的唯一功能性意义，尤其是在大宗商品领域，就是为了套期保值（hedging）。投机当然也很好，但核心价值是让企业管理波动性、进行风险转移与资本配置；如果我们做出来的指数连对冲都做不到，那我们的工作就是彻底失败的。 — Carmen Li > `13:30` 如果价格注定永远上涨，我们干嘛还需要指数？ — Carmen Li > `13:30` 这真是一件极其迷人的事情，这就是我每天要回答的所有问题。 — Carmen Li > `20:00` 市场永远是用脚投票的（the market always votes）。 — Carmen Li > `20:00` 出于私心，交易的人越多对我自然越好。 — Carmen Li > `30:00` 我开自己的车开了快十年，汽车的折旧贬值幅度可比这个夸张多了！ — Carmen Li > `30:00` 我不知道市场上所谓‘两年就彻底报废贬值’的数据是从哪里编出来的，如果有人愿意在第二年以一折（10 cents on a dollar）的价格把二手芯片甩卖给我，我愿意每天都全盘接下这笔交易。 — Carmen Li > `32:20` 所有的星象都必须完全对齐（all stars have to align）。 — Carmen Li ## Mentioned Carmen Li, Don Wilson, Tracy Alloway, Joe Weisenthal, Michael Burry, Silicon Data, Compute Exchange, DRW, CME, CFTC, Bloomberg Terminal, Jefferson Lab, Platts, CARFAX, Polymarket, Carvana, OpenAI, Anthropic, Google, SpaceX, Jane Street, Shell, American Airlines, Nvidia A100, Nvidia H100, Nvidia B200, Nvidia B300, Nvidia L40S, Cerebras, SiliconMark ## Video description <p data-renderer-start-pos="94">When we spoke to DRW's Don Wilson last year, he talked about building out a GPU market that might be bigger than oil. Now, a year later, he is working with Carmen Li to do just that. Li is the CEO of two companies &mdash; Silicon Data and Compute Exchange (where she works alongside Wilson). The former company is building the index for GPU pricing while the latter is a spot marketplace for GPU procurement. Today's episode &mdash; recorded at our live show at City Winery in New York &mdash; gets into how Li is building a whole new market for GPUs at her two companies. We talk about the challenge of standardizing compute, GPU price volatility, if used GPUs are like used cars, what goes into constructing a GPU index, and what it means to win the GPU lottery.</p> <p data-renderer-start-pos="842">Read more:<br><a href="https://www.bloomberg.com/news/articles/2026-06-04/jane-street-plans-new-data-center-as-compute-power-runs-scarce?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Jane Street Plans New Data Center as Computing Power Runs Scarce</a><br><a href="https://www.bloomberg.com/news/articles/2026-06-05/google-buying-computing-from-spacex-in-920-million-a-month-deal?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">SpaceX Inks $30 Billion Computing Power Deal With Google</a></p> <p data-renderer-start-pos="976">Only <a href="http://bloomberg.com/" data-testid="inline-card-resolved-view">Bloomberg - Business News, Stock Markets, Finance, Breaking & World News</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at&nbsp; <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">bloomberg.com/subscriptions/oddlots</a></p> <p data-renderer-start-pos="1148"><a href="http://bloomberg.com/subscriptions/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Subscribe to the Odd Lots Newsletter</a><br><strong data-renderer-mark="true">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p> [Source](https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/8a94442e-5a74-4fa2-8b8d-ae27003a8d6b/60472ab3-350e-4e47-a612-b4670121babd/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[10:00] Carmen Li: 作为最终用户，你可以明确决定价格是否是唯一的首要考量；或者，你是否心甘情愿为了更好的地理节点低延迟、更强的数据吞吐性能而支付溢价。 [11:15] Carmen Li: 金融产品存在的唯一功能性意义，尤其是在大宗商品领域，就是为了套期保值（hedging）。投机当然也很好，但核心价值是让企业管理波动性、进行风险转移与资本配置；如果我们做出来的指数连对冲都做不到，那我们的工作就是彻底失败的。 [13:30] Carmen Li: 如果价格注定永远上涨，我们干嘛还需要指数？ [13:30] Carmen Li: 这真是一件极其迷人的事情，这就是我每天要回答的所有问题。 [20:00] Carmen Li: 市场永远是用脚投票的（the market always votes）。 [20:00] Carmen Li: 出于私心，交易的人越多对我自然越好。 [30:00] Carmen Li: 我开自己的车开了快十年，汽车的折旧贬值幅度可比这个夸张多了！ [30:00] Carmen Li: 我不知道市场上所谓‘两年就彻底报废贬值’的数据是从哪里编出来的，如果有人愿意在第二年以一折（10 cents on a dollar）的价格把二手芯片甩卖给我，我愿意每天都全盘接下这笔交易。 [32:20] Carmen Li: 所有的星象都必须完全对齐（all stars have to align）。

[0:00] Tracy Alloway: compute is the new oil [9:07] Carmen Li: there is up to a 38% performance variance among identical chip models. [15:00] Carmen Li: what's in it for me? [15:00] Carmen Li: hey, you know, if I give you all my data, I give away all my secrets. [20:00] Carmen Li: the market always votes [22:45] Carmen Li: The beauty of CME is you can do the same thing you're doing now, treating CME products... Everything is BAU. It's no different. [28:40] Carmen Li: The second year H100 residual value, resale value for refurbished chips about 85 cents on a dollar... The third year is 84 cents on a dollar. I think my car depreciated way more than that.

[0:00] Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget, with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. The thing about AI for business? It may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slash repetitive tasks, and freed thousands of hours for strategic work. Now we're helping companies get smarter by putting AI where it actually pays off, deep in the work that moves the business. Let's create smarter business, IBM. When you're running a business, The best days are the ones where priorities stay on track. For midsize and large companies, that isn't always easy. Risk can touch multiple parts of an organization at the same time often in ways that aren immediately obvious It might involve property liability or cyber It could stem from regulatory requirements or challenges tied to a specific industry or the scale of an operation At that level managing risk becomes an ongoing discipline not a one decision At The Hartford, the focus is on helping businesses manage risk before it turns into something more disruptive. That means working with companies to identify where they're exposed, decide what matters most, and put practical standards in place so risk is managed as part of day-to-day operations. And when losses do happen, The Hartford can pair that risk control work with insurance coverage grounded in underwriting, risk engineering, and claims experience developed over time. Learn more at the Hartford dot com slash risk mitigation. Bloomberg Audio Studios, podcasts, radio news. Hello and welcome to another episode of the All Thoughts podcast. I'm Tracy Alloway. And I'm Joe Wasenthal. So, Joe, we're still continuing our series recorded from the live show in New York. We had a bunch of great conversations A couple of them were building off of discussions that we had had previously And one of those discussions was in Chicago at another live show about six or seven months ago, back in October, we spoke with Don Wilson of DRW about the trading environment, but also about his new venture. Right. And so his new venture is one that actually there's quite a bit of competition and quite an excitement. And it's essentially like, OK, GPUs, we know they're very important for the AI boom, et cetera. The question is, can GPU capacity, which is scarce, can it become a tradable commodity such that I can buy futures to lock in my price of access to compute power? Could I resell those futures? Will there be speculators speculating on the up or down price of like an H100, running an H100 NVIDIA chip for an hour? This is a big question. We know there's a lot of interest in the actual compute, but whether there's interest in compute futures as tradable instruments is very TBD. Yeah. And the analogy that everyone always uses is compute is the new oil. So why can't it have, you know, a market structure that looks somewhat like the oil market? And there are challenges Fungibility is a big one Like one chip might not necessarily be equal to another chip Or one chip the same chip at one data center might not be equal to the same chip at a different data center Exactly And so even if you not interested in A what I say here is like the market structure questions and the idea of building an entirely new market is really fascinating to me And I think others will find it interesting, too. And we really do have the perfect guest. We're speaking with Carmen Lee. She is the CEO of Compute Exchange and Silicon Data. These are the two companies that Wilson is invested in. And they've already announced that they're doing futures with the CME. So really the perfect person to speak to. So take a listen. Last October, we spoke with Don Wilson of DRW fame, and he was talking to us about his new project, which was basically building out this Compute Exchange. Now we're here with you six months later. you're actually the one leading it. How far are you in this endeavor? And remind us, what exactly are you trying to do here? Yeah, so thank you for the great intro and great audience. Before I do that, I'm actually going to call back to six months ago in the Don podcast you did. You asked me the question, what if compute prices keep going to go up? At that [5:00] September, October, compute prices were going down across all chips. Now see what happened. I think you called it. I think you called it when the market called it. So I'm the founder CEO for Silicon Data. So that's the index provider for GPU indices. We recently announced partnership with CME. So we will be launching GPU future options on CME in a couple of months, pending SAP TC approval, obviously. So that's quite exciting. We've been working on GPU indices for the past two and a half years, starting 2024, April. So it's been a while. And we launched world's first GPU indices at Bloomberg Terminal in 2025. A year later, we launched the partnership with CME. So it's quite exciting. Separately, I heard you mention Compute Exchange before. So thank you for doing that, Joel. I'm the CEO for Compute Exchange, which is spot marketplace for GPU procurement. So we do reserve contracts, forward contracts, as well as refurbished contracts. Let's talk about the variety of options that we have to financialize compute and so forth. So this came up in the first conversation we had with Ian Dunning Who is the type of buyer who would want to buy compute on a spot market Because right you talk about typically we think it like these multi contracts where some entity enters into a contract with a data center or a new cloud, whatever, and they have this for a while. So who is the buyer or the user of these instruments that might want to buy spot compute or very short-term, short-dated compute futures? It's a great question. So the compute market right now for Compute Exchange, we have all our provider mostly are NeoClouds around the world. That's one side. Another side is a big variety from AI startup. So even though they are startup, they spend millions of dollars on GPUs already. There are enterprises who are traditional businesses, but they are needing a node, two nodes, a few servers here and there for their inferencing or, I don't know, other deployment needs. They are providers. They are inferencing providers, right? They don't own GPUs, but they provide open source, open weights, model support for other use cases. So what's the big variety? Most North American firms they do a variety of combination of contracts Obviously on give you the most flexibility You don pay when you don use it However you also at the mercy of demand supply curve at any given time So translate to your price can go from $3 to $6 to $9, depends on demand supply curve shifting. So that doesn't help when you can have a predictable margin. And also, in terms of scarce, you're not guaranteed for your GPU resources for next hour or next month. So you see a lot of people shifting from on-demand to reserve even forward contracts, right? So forward contracts, you basically lock in deliverables for next whatever month, right? Starting in September, maybe, or starting November, February opens, right? So this all comes because of market condition. So Compute Change cover that, the physical GPU procurement, also token. So we'll talk about token as well. On flip side, who's going to use the futures options can be a similar set of people, right? You look at oil market, which we all love WTM brands, right? The people use WTM brand, a lot of them are naturally long oil. So the shells, right? The producers, they need to hedge their revenue volatility by shorting futures or port options If you naturally short oil American line right They want to control their cost volatility They want to obviously use future options as well Simple to compute. Your new cloud or you or anyone have the servers. Ideally, you want to have predictable revenue streams. So the new cloud would be the shell in this example. Exactly. You have GPUs, right? Or the banks where GPUs own your balance sheet, right? Your long GPUs. then naturally you want to make sure revenue, right, is stable to a certain degree, and then you want to use future to do so. If you are naturally short GPU, which is everybody in this room, unless you tell them you have GPUs, right? Then it depends how much you use. If you want to control your cost volatility, you want to use future to hedge as well. Just on the compute exchange side of things, if someone is buying like off the spot market, how do you guarantee, I'm not sure quality is the right word for this, but how do you guarantee they're getting what they expect? This is a great question. So I'm going to flip to slide if you don't mind. Oh yeah, we have visuals for this one too. So I usually don't like to use slides, but this time, because you mentioned really good questions, so we actually call it GPU lottery. So we published a paper early this year at GPG, [10:00] conference with Jefferson Lab on GPU performances. We can have you create a link to the audience later on. This is A100, by the way. I know we didn't put a tag on. This is A100, 40 gigabytes memory bandwidth. We proved there's 38% performance variance for the same chip, and then we decomposed into the chip itself, intra-provider, and inter-provider. And there's many reasons for that, right? And to your point, you don't know until you get your GPUs. We have a PLATS for GPU, CARFAX for GPU, depends how you look at it. So in compute change, you actually verify the GPU before it's delivered to you. So basically, you can RPEU for, say, hey, I want 200 B200 nodes. Obviously, we'll give you specs back and the commercial back. Same time, independently verify the performances on flops, memory bandwidth, tokens, and other information, SLAs and other things. And as a user, you can decide, is price your most important criteria? Maybe it is. Or maybe you're willing to pay a premium for geolocation or the performances that you care more about on latency right We believe give people the option and transparency is the most important thing Let stick with the oil analogy for a second You know there a few benchmarks that we all know about There's Brent, there's WTI. There's others, but those are the two that we talk about. If we transpose this to chips for a second, okay, let's say you have an H100 index. We did an episode of the podcast last week, I think, with the CEO of Cerberus, which is another... It's an amazing company, yep. Yeah, but they're another type of chip for inference. Is your assumption that these indices are going to be close enough to the cost such that if you're, okay, I'm running inference maybe on some servers or TPUs or whatever, some of these others, that an H100 index will be good enough as a hedging instrument? This is the whole goal for me sitting here, actually, right? There's a meaning for every financial products. In the functional reason, for commodity, it is for hedging. Right? This speculation is great, but it's really for people to hedge their volatility, to do risk allocation, to do risk transfer, and then asset capital allocation. If we can do what you said then we fail at our job Right So that why we went all the way back The way we developed our index model is not a simple math It not hey you have 2H100 to a simple average right Because then you compare Apple to oranges. The 2H100 can have different CPU, different RAM, different disk, different dual location, different memory bandwidth. You cannot do simple math. What we do is we usually collect six months of historical trading data from over 100 data sources. and we see which factor drive the price differentiation. So every day, over 150,000 traded prices ingest in our platform, and we normalize the traded prices based on different characteristics of the model itself and normalize to a base case, and then we do the math of settlement price calculation. So then this price will be highly correlated, ideally, as much as it can, to the price you pay at a new cloud, for example. However, it won't be the same, just like basis trading. Like every other commodity, there's a basis risk. We're helping clients calculating the basis risk. So you know, hey, you're US East, you may be a bit higher or two, then there expectation a manageable correlation understanding of the indices You mentioned volatility just then I mean the reason people need to hedge is because of volatility Are you seeing enough of that in GPU prices that like this model makes sense? Because if it's just a steady line up or steady line down, like it's going to be a kind of boring market. So it's interesting. So last year when GPU price all going down, the big conversation is why do you need indices for something price will always go down. And this year, why do you want indices when prices always go up? Literally, that's all the questions I get. It's pretty fascinating. So when we look at volatility, we look at daily volatility movement, not the price up and down. The daily volatility for A100, H100 is around 20 to 30. It's a very healthy commodity volatility range. So I don't manage volatility. It just happened to be that volatility. That can't change. It's all because we normalize it. If you look at each individual chip configuration at different geolocation, the volatility are different. There are some chips with 8% volatility, some chips with over 100. Because normalization of indices, you actually get very healthy 20 or 30 daily law. [15:00] Thank you. Thank you. The show comes from public.com. If you're actively involved in your portfolio, you probably catch yourself repeating the same actions, buying the dip manually sweeping idle cash putting on a hedge On public you can now create AI agents that handle all these tasks on your behalf Just describe what you want to do in plain English Like if the VIX hits 25 buy a put option on the S 500 Or, if my cash balance goes above $20,000, move the excess into my direct index. You approve of the workflow, and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., member FINRA and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. The thing about AI for business, it may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slash repetitive tasks and freed thousands of hours for strategic work Now we helping companies get smarter by putting AI where it actually pays off deep in the work that moves the business Let's create smarter business, IBM. I'm always fascinated by – we look at the Bloomberg terminal, for example, and there's a price on the screen, and it's just there. And we sort of take it for granted that it had to come from somewhere. and maybe some commodities have like a, you know, there's an existing exchange and a public price, and then there's also a lot of commodities, just bilateral trades. What is the actual process by which you collect the most recent data? So if you say, okay, an hour of H100 usage costs X, right, whatever it is right now, how did you assemble that number? How did you gather that information from say the inference providers so it is a very can be lengthy depends on what data sources the nature of gpu spot markets so compute changes one of them and then many many new clouds hyperscaler marketplaces all have very different contract size um duration specs and their way to manage their data right so it a lot of a licensing conversation negotiation a good well and also context myself I don know I was used for Bloomberg data So I was in data business for a period of time So everything is pretty intuitive to me It's very important to get a variety of data sources, especially for computing. But you call them up. So it's like, OK, the price is different on a Friday versus one. You zoom them up. Yeah. Well, first you have a conversation. Say, hey, I love what you do. You're a great new cloud. Can I license your data? OK. And usually your feedback is what is in for me. Yeah. Right. And then we'll talk about commercials. And then your concern could be, hey, you know, if I give you all my data, I give away all my secrets. And then we'll go through traditional licensing agreement. Can I disclose what I want from you? What I do not want from you? What's the pipeline look like? Are you writing a straight bucket job? Are you writing my API to yours? Are you writing to mine? It's a lot of conversation. It's actually pretty standard conversation. And right now we have 8 million pricing points globally around 200 data sources. It's pretty much BAU. A lot of people will say, hey, karma always bring up, can I have your data? It has always been ending. You know, we were talking about GPU indices and you're not the only one doing GPU price indices for sure. Not anymore. Yeah, not anymore. But when you look at some of the other ones, like sometimes they show different, [20:00] numbers or even different longer-term trends. What accounts for the discrepancy there? What are you doing differently, or what are they doing differently, I guess? So I can't comment on other people's mythology because I actually don't know. Different raw data, different mythology will eventually drive different prices. So the way I would look at this is, you know, it's always smart for anyone to look at multiple data sources and then figure out what is the actual decision you have to make, which data source do you trust, the market always votes. Once things start trading, the market always gravitates with things that actually help them hatch. If you're easily manipulatable, if you are not data source, people actually do you hatch. What's the point, aside from speculation? So I love to say, I mean, I also strongly believe we're the best, But again, I will let the market decide, which will happen very soon. So of course, yes, there's the economic rationale for the existence of a hedging instrument. And we can understand that Someone who is an entity that needs compute they implicitly short GPUs they want to hedge et cetera But the liquid markets also really do need speculators and they need people betting on price What are you seeing right now in terms of traders or institutions, etc., who economically can take both sides of the trade? And how active is this getting where there's just a compute trading desk that is separate from their economic needs? The conversation has been going on for a very long time with various banks, various market participants, speculators they're all very excited so some banks obviously have both sides of the trade, right? So they can cross off some positions internally, that's great obviously some they have to use leverage external products that's where we come in. The way I encourage them to do is selfishly I want them to start trading that's going to compute. The more people trade the better for me, right? Selfishly but at the same time it's important for people to understand GPU trading It's not like you can't just move someone from all your electricity with no background context drop into GPU compute features. There's a lot of context where number one, GPU, it is not a homogeneous product. Number two, you have to understand the use cases of A100, H100. Right now, they're not that correlated. Is that right Maybe that not right I don know There are use cases which they pretty separated but maybe their use cases they can be transferred And also there software layer to this right So right now you can argue certain use cases some large models cannot be deployed and the legacy chips, but doesn't mean six months later you cannot do so. As the software layer compression, model compression gets better, optimization gets better, things can change. So really understand not just the hardware configuration, this local supply demand curve for the server itself, also the software layer. That's kind of critical, right? That's really changing the supply demand curve and all the way to the user behavior. So it's going to take some time since we have engaged with a lot of participants. Make sure they have the right setup. I have what is possibly a dumb question, but the compute futures, how are those actually settled? Because I have images in my mind of taking physical delivery of maybe one of those big... I'll give you a wafer. I'll give you a server wafer. That'll be fun. So for the CME futures, we'll be financially settled, just like the traditional oil settlement prices goes. Four contracts, obviously we do four right now at Compute Exchange, but we're always open to do, you know, physically delivered futures, especially given we do have Silicon Mark which is GPU benchmarking So imagine in the future you can do hey I want a 20 grade A B200 this configuration this shape of servers in U East And then at the end, we'll get that. Well, usually API calls, so you don't get physical. It's not as cool as physical to give you a way for. But you get API calls. One can dream. How do you literally trade it? Is in like, let's say there's probably some very bright people in the room, not with an institution. When it's all listed and everything, does it need to go through like a futures broker? Is it like, could it be like a prediction market where you just go to the website? Like what is the actual, how does someone actually get in this? Setting aside whether they're sophisticated enough and whether they know what they're doing. A lot of people in trade who have no idea what they're doing. Yeah, setting all this aside, yes, you know, only trade what you know. But like, what is it through a prime broker? or how will people actually be able to participate in this market? The beauty of CME is you can do the same thing you're doing now, treating CME products. Okay. The same process, other processing margin, that's why you get great margin optimization, right? Everything is BAU. It's no different. We don't have anything right now. So any commodities broker that someone has, they will be able to, on that platform, they will [25:00] access to these instruments exactly right yeah we make it easy for people would you be upset if a prediction market set up a gpu price contract of some sort with that into your business not at all um so we actually work with polymarket uh last year someone actually listed my product at polymarket without my consent it's pretty it's always start like that and then someone told me that and then we tried to polymarket say hey do you want to do something you know more real so we We did February settled and April settled a few contracts on Polymarket just to test the water, right? Obviously, we're exclusively with CME right now. But yeah, so I think obviously you have to do the right licensing, nominal pilot, right? All the right things. Yeah, you know, I don't, Mark can do whatever they want. And then people will choose the best product for them to use. Setting aside the financial instruments for the moment, would people think about AI and they think about the use of GPUs, they mostly still probably in their mind think of like OpenAI, Anthropic, and Google basically. And that's kind of it. But obviously as you stated the world of entities that serve inference in some form or another is much greater than these three companies that we talked about Talk to us a little bit more about what the actual world of inference provision looks like outside of the big household AI names. So the ones you mentioned, they mostly are closed source models, as we call it, right? But they do have some open source versions, but they're famous for their closed source models. So we actually track 300 open source, open weights, closed source models globally from pricing and consumption point of view. It is really interesting if we have actually, you know, we haven't really formally launched LM token indices. You can kind of look at Bloomberg and it's on Bloomberg. What's interesting is people are, depends, it's all based on your choices. Right now, the price actually doubled from our indices from now to December 1st last year. It's like $2.21 per million token. It's a mixture of input-open token prices, average weighted by consumption by bus-scale models. It's not here. This is GPU, unfortunately. Wait sir Since we have this specific chart up right now what is the y in this chart Sean So we looking at the dot per GPU per hour rental rate on demand for three chips The top one the yellow line is B200 NeoCloud on demand per GPU per hour Sorry, it's a mouthful. The line, the yellow line is interesting, right? So every new chip came out, based on historical data, A100, H100 usually came out to be high. And then comes down as more supply came live. and then price came down and then stabilizes. So that's the trend we have observed for A100 and then for H100. So when B2 country came out, we published the data last year at Bloomberg, this early this year, the price was high and it came down, which is kind of what I expected, but the slope was less steep than I expected. I was like, hmm, that's interesting. The slope wasn't as steep. And then quickly observed the price just came up. And now it's higher than the initial open, whatever you call that, right? launch prices. That shows you demand supply curve in a different stage than whatever stage we had before. So the red line is H100, NeoCloud on demand per GPU per hour rate. So you can see the price came down last year a little bit. Sorry about the scale so you don see much but it came down and came back up quite a bit I think the last three months came up to like 8 for the H100 The H100 is the oldest chips among the three right They're pretty, you know, pretty much a commodity at this point. The price came down, they stabilized, but the price came up about 10, 15% for the past three months. Remember, the H100, right? They're not the latest and greatest at all. So this also tells you the supply demand curve shifting. Oh, yeah. Actually, that reminds me. Can you talk to us? Because you're doing refurbishment of chips as well, right? Which seems challenging in many ways and kind of reminds me a lot about the sort of Carvana model of compute or something like that. How are you actually doing this? How does that business work? So this is calling two different things. One is for people come to Compute Exchange saying that, hey, I want to, you know, it's a new cloud provider, right? If you get a piece of land, you get energy, a co-location, great, congratulations. Then your option is, number one, should I get the latest and greatest? The B300, the GBs, the Verubon wait for a few months, or do you want to get refurbished trips and turn it on maybe sooner, right? Then to you, it's become ROI calculation for the most part, right? What's your expected... [30:00] You know, future revenue generation, what's your residual value calculation, how much you can purchase by, right? It's actually pretty simple cash flow based ROI calculation. So the way we approach residual value and the refurb of transaction is, you know, based on ROI, you know, this is your potential break even. Look at H100, right? Obviously, you're not going to charge as high speed 200, but your cost base is also lower. So you can do the future, as you assume, a few years of forward contract, you're signing three years, discount the cash flow back, that's your residual value now, right? So we do that calculation with people so they understand, hey, what's the value supposed to generate? And then what's the trading, the market prices for refurbish or use GPU? Again, you have to test you to make sure things work and there's other nuances to that. But we help with the understanding of the whole residual value. And that's why the whole bubble thing came about. But go ahead. What month was it last year when everyone got really obsessed with the lifespan of chips? November, December. Remember Michael Burry tweeted something about, he's like, oh, the lifespan, they're like, right? And everyone just spent like three weeks free and then moved on from that conversation right Like that was like what do we know about chip lifespans are there misconceptions out there about the how long yes these can get be productive i got interviewed a few times but i don't i mean i'm not important i still am not important but back then i even less relevant i was telling reporters i was like look i don't know what data you're looking at um based on my i actually have blogs so my website which is completely you can just search for it last year because of that conversation. I want you to curve later on. The second year H100 residual value, resale value for refurbished chips about 85 cents on a dollar. So a year later, you can sell 85 cents on a dollar. That's pretty good, I would say. The third year is 84 cents on a dollar. I think my car depreciated way more than that, right? And I drive my car for whatever, 10, 10 years. So I had the data, but again, I'm not going to argue against narrative, which is... So, but there's a fairly seen, a decent drop from year one to year two but after that you see a general level that's november december analysis right now it's a little different i haven't refreshed the study but our code is there if you're my data client if you're wrong my code you get a number right away the another thing i want to point out is l40s they're like you know the ogs right at that time these years people still use them they charge you hyperscalia charge you 40 cents per gpu per hour So you know I don know about two years where that number coming from But I will I will do that trade every single day You sell me your two years old at 10 cents a dollar I will buy it. There's a sort of big question looming in the background of a lot of these discussions, which is the B question, I guess. Whether or not we're in an AI bubble. Right. And you sort of touched on it earlier. You have all this granular data on how people are actually using compute, GPU prices, all of that. What's your take on the big question? So as an index provider, I cannot give any forward guidance. Not a disclaimer. Nor do I know, right? In fairness, I know what do I know. So the way I look at it is we have defined a bubble, right? So you look at a stock bubble, right? And the NASDAQ showed up 200%, it came back down to 84%, whatever, back then. it's a bubble, right? The way I look at bubble is, is your valuation, can your future cash flow support today's valuation of yours, right? So then I'm not talking about like opening an everyone else valuation. I'm not VC. I don't understand that process. The way I look at GPUs the machines it actually pretty simple Look at future cash flow of your forward contracts and then you discount it back Can you get your money back for the price you pay It's actually pretty straightforward for the machine level. But to your point, you can say, hey, what happened if demand dropped? No one's going to use your whatever things you have. But remember, the forward contract is a signed contract. If you have that, you kind of know. Obviously, if you have things, the biggest concern is people have concern overbuilt. If you've overbuilt, then by theory, then all your prices will calm down because of the oversupply of the market, right? So then you talk about supply-demand equilibrium. How do we know about future demand of GPUs, right? I don't know that. Everyone's guess is better than mine, probably. the way I look at it, it's not that easy to bring any GPU online, right? You hear all those say big data center bill, $25 billion invested, but they don't translate to immediate GPU availability, right? In the servers, which you have to be a way listed, if you buy brand new stuff, co-location, you need optic fiber. So it's a lot of, unfortunately, a star has to be [35:00] Wait, but people can default on contracts, right? So even if you have a long-term contract signed, that could not work out. Could you envision credit default swaps or something in the compute market? So that happens in every other market. Every market, if you do OTC trade, you have to raise someone who will default you. It doesn't matter who they are. So there's a lot of mechanism to hatch that. The things you cannot hatch is GPU cost, the price you entered. So that's something exactly what CME Futures is for. You can have the transparency, the liquidity, and then the easiness of trading in and out and hedge your position. Carmen Lee, thank you so much for joining us. Yeah, thank you. That's great. Thank you That was our conversation with Carmen Lee of Compute Exchange and Silicon Data recorded live at our New York show I Tracy Allaway You can follow me at Tracy Allaway And I'm Joe Weisenthal. You can follow me at The Stalwart. Follow our guest, Carmen Lee, at Carmen Lee. Follow our producers, Carmen Rodriguez, at Carmen Armin. Dashiell Bennett, at Dashpot. Kale Brooks, at Kale Brooks. And Kevin Lozano, at Kevin Lloyd Lozano. And for more OddLots content, go to Bloomberg.com slash OddLots. We have a daily newsletter and all of our episodes. And you can chat about all these topics 24-7 in our Discord, discord.gg slash OddLots. And if you enjoy OddLots, if you like it when we do these live shows and talk about the birth of a new market, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad-free. All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening. Music Thank you. for your brand and budget with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. When you're running a business, the best days are the ones where priorities stay on track. For midsize and large companies risk can affect multiple parts of the organization at once from property and liability to cyber and regulatory challenges At that level managing risk becomes an ongoing discipline. At the Hartford, the focus is on helping businesses manage risk before it turns into something more disruptive. And when losses do happen, that work is paired with insurance coverage shaped by years of underwriting, risk engineering, and claims experience. Learn more at thehartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. Wise is the smart way to manage the currencies you need around the globe. When you send money abroad using your bank, you could get hit with hidden fees and exchange rate markups. There's a better way. Try Wise. Wise uses the exchange rate you'd usually find on Google with no unwelcome surprises. Plus, most transfers happen in under 20 seconds, which means your money arrives in less time than you've been listening to me. It's simple and free to sign up when you download the WISE app. Be smart. Get wise. T's and C's apply.