人工智能正在改变现代世界的许多事物。语言也不例外。例如，“compute”（计算）一词过去被视为动词。如今，即使是语法学究也勉强接受它作为一个名词，指由先进芯片产生的处理能力。这种新型的、非动词的“compute”（算力）是一场技术革命的燃料。
ARTIFICIAL INTELLIGENCE is changing many things about the modern world. Language is no exception. The word “compute”, for instance, used to be regarded as a verb. Today even grammar pedants grudgingly accept it as a noun, referring to the processing power generated by advanced chips. This new, non-verbal compute is the fuel of a technological revolution.
对处理劲头（算力）的需求激增，且供应短缺。根据波士顿咨询公司的统计，去年全球在人工智能算力上的花费约为3600亿美元，比铜多出约50%。波士顿咨询公司预计，到2030年，相关交易额将达到2.3万亿美元，大致相当于当今全球原油贸易的规模。
Demand for processing oomph has surged, and it is in short supply. About $360bn was spent on AI compute last year around the world, according to Boston Consulting Group, about 50% more than on copper. By 2030 BCG expects $2.3trn in transactions, roughly the size of the global trade in crude oil today.
然而，尽管算力市场规模惊人，却缺乏成熟度。“超大规模计算商”（这是一个对少数几家巨型云计算服务提供商（包括亚马逊、谷歌和微软）而言仍然有些晦涩的IT行话）以及前沿人工智能实验室（如Anthropic和OpenAI）达成价值数百亿甚至数千亿美元的算力交易。但成千上万家小型企业却无法获得此类安排。买家和卖家也几乎没有衍生品，而在成熟的市场中，衍生品能让他们对冲风险并进行投机。
For all its impressive size, however, the compute market lacks sophistication. The “hyperscalers” (a still somewhat arcane piece of IT jargon for a handful of giant cloud-computing providers including Amazon, Google and Microsoft) and frontier AI labs (such as Anthropic and OpenAI) make deals for tens or even hundreds of billions of dollars of compute. But such arrangements are unavailable to thousands of smaller firms. Buyers and sellers also have few of the derivatives that in established markets allow them to hedge their exposures and speculate.
因此，算力供应链看起来正适合市场化。最简单的形式是，芯片制造商（如英伟达和AMD）将先进的人工智能半导体出售给设备制造商（如戴尔和慧与科技），这些制造商将硅片组装成机架和服务器，出售给数据中心运营商。除了超大规模计算商之外，这些运营商还包括少数大型“新云服务商”（又一个语言创新，指CoreWeave和Nebius等独立提供商）以及数百家规模较小、知名度较低的公司。
The compute supply chain, then, looks ripe for marketisation. In the most simple form, chipmakers (like Nvidia and AMD) sell advanced AI semiconductors to equipment manufacturers (such as Dell and Hewlett Packard Enterprises), which package the silicon into racks and servers to sell to data-centre operators. Besides the hyperscalers, these include a few large “neoclouds” (one more linguistic innovation, denoting independent providers like CoreWeave and Nebius) and hundreds of smaller, less well-known firms.
反过来，运营商则在固定期限（通常为几年）内将算力的访问权限租给客户。客户范围从超大规模计算商（它们用租赁来补充内部算力）到运行开源人工智能模型的初创公司。
The operators, in turn, rent out access to compute to customers for a fixed period, often several years. Customers range from the hyperscalers (which supplement their in-house compute with rentals) to startups running open-source AI models.
到目前为止，理论上很简单。在实践中，情况很快就会变得复杂。这是因为计算能力是异构的：不同的芯片在不同的服务器上销售，位于不同的地点，时间期限也不同。匹配买家和卖家是很棘手的事情。每个需要处理能力的企业都有一个“计算负责人”（或数人），通过不透明的中介网络购买二手算力。许多这类非正式经纪人，有些比其他人更可疑，通过 WhatsApp 和 Telegram 等即时通讯应用推销他们的产品。整个过程看起来不像是交易高科技商品，倒更像是贩卖非法毒品。
So far, so simple—in theory. In practice, things quickly get complicated. That is because compute is heterogeneous: different chips are sold in different servers, at different locations, for different time periods. Matching buyers and sellers is finicky. Every business that needs processing power has a “compute guy” (or several) buying second-hand capacity through an opaque network of middlemen. Many of these informal brokers, some shadier than others, tout their offerings through messaging apps like WhatsApp and Telegram. The whole thing looks less like trading a high-tech commodity and more like peddling illegal drugs.
在这样一个支离破碎且不成熟的市场中，供应商和买家都更倾向于长期合同带来的安全感。从纸面上看，这些合同向用户保证了算力，也为供应商保证了收入。在现实中，鉴于人工智能技术和市场的发展速度如此之快，很难判断此类交易的价值究竟有多大：今天的买家和卖家在几年后可能不复存在，无法履行其承诺。
In a market that is so fractured and immature, providers and purchasers alike prefer the safety of long-term contracts. On paper, these guarantee capacity to users and revenues for themselves. In reality, given how fast the technology and the market of AI are evolving, it is hard to tell how much such deals are worth: today’s buyers and sellers may not be around in a few years to honour their commitments.
CoreWeave 债券收益率的飙升突显了这种不确定性。谁会赚得盆满钵满？而谁又可能根本无法存活？由于贷款机构难以对风险进行定价，许多购买和租赁算力的小型运营商面临着两位数的利率。
The surging yields on CoreWeave’s bonds highlight the uncertainty. Who will make a killing? And who may not make it at all? With lenders struggling to price risk, many smaller operators buying and renting compute face double-digit interest rates.
对所有相关人员来说幸运的是，围绕算力供应链的正规市场正在建立。一些新兴的交易平台，例如 Compute Exchange，按地点、租赁期限、价格和可用性列出 GPU（图形处理单元的日益熟悉的简称，即人工智能的硅基主力）。全球衍生品交易巨头，包括芝加哥商品交易所集团和洲际交易所，希望推出自己的算力期货。它们已与开创算力指数的初创公司合作，这些指数与热门人工智能芯片的租赁价格挂牌。它们正在等待监管部门的批准。
Fortunately for everyone involved, proper markets are being built around the compute supply chain. Some budding bourses, such as Compute Exchange, list GPUs (the increasingly familiar shorthand for graphics-processing units, AI’s silicon workhorse) by location, rental duration, price and availability. Titans of global derivatives trading, including CME Group and the Intercontinental Exchange, hope to launch their own compute futures. They have teamed up with startups pioneering compute indices, pegged to the rental price of popular AI chips. They await regulatory approval.
某些算力市场的先驱并没有等待交易所交易衍生品获得批准，而是在创建类似于信贷和外汇市场上常见的场外交易工具。今年3月，一家名为阿内拉实验室（Anera Labs）的初创公司与加密货币经纪商法尔康交易所（FalconX）安排了一笔六位数金额的试点交易，交割时间为6月份，标的为英伟达广受欢迎的H100图形处理器（GPU）的未来租赁价格。
Rather than wait for exchange-traded derivatives to be approved, some compute-market pioneers are creating over-the-counter instruments of the sort common in credit and foreign-exchange markets. In March a startup called Anera Labs arranged a six-figure pilot trade for June with FalconX, a cryptocurrency broker, based on the future rental price of Nvidia’s popular H100 GPUs.
另一家初创公司卢克索（Luxor）向运营商提供类似互换（swap）的合同，保证每台服务器每小时的固定现金流。如果算力的现货价格低于互换协议中商定的价格，运营商将获得差额赔付。如果现货价格更高，则由运营商支付。该公司的联合创始人之一伊森·维拉（Ethan Vera）解释说，这笔支付可以通过收入分享安排等方式进行。卢克索的定制衍生品目前规模还不大，在数千万美元左右，但潜力巨大。
Luxor, another startup, offers operators swap-like contracts that guarantee fixed cashflow per server, per hour. If the spot price for compute is below the price agreed in the swap, the operator is paid the difference. If the spot price is higher, it is the operator that pays. This payment, explains Ethan Vera, one of the firm’s co-founders, may for example be made through a revenue-sharing arrangement. Luxor’s bespoke derivatives are modest for now, in the tens of millions of dollars, but the potential is vast.
正当美国各大交易所等待监管机构的批准之际，新兴企业正在其他地方起步。成立于2023年的架构金融技术公司（Architect Financial Technologies）在其位于百慕大的交易所提供与美国芯片价值挂钩的算力期货。该公司最近还推出了一个平台，供模型制造商竞标提供“推理”（即人工智能模型对提示词作出响应而产生的有用输出），而不是底层的算力容量。该公司创始人兼老板布雷特·哈里森（Brett Harrison）认为，事实证明，“推理”可能比算力更容易标准化。
As American exchanges await the thumbs-up from regulators, upstarts are getting off the ground elsewhere. From its exchange in Bermuda, Architect Financial Technologies, founded in 2023, offers futures in compute linked to the value of chips in America. The firm has also recently launched a platform where model-makers bid to provide “inference”—the useful output produced by an AI model in response to a prompt—rather than the underlying compute capacity. Brett Harrison, the company’s founder and boss, believes that inference may prove easier to standardise than compute.
鉴于这一全新市场充满不确定性，它正吸引着风险承担者。许多人（包括哈里森）曾在加密货币领域摸爬滚打，那曾是算力需求最旺盛的行业，吸引了大量金融工程师试图将算力转化为可交易资产。哈里森曾任FTX US总裁，该公司是山姆·班克曼-弗里德创立的加密货币交易所的美国分支机构；他在2022年11月母公司因欺诈破产前两个月辞职。
Given the uncertainty surrounding the all-new market, it is attracting risk-takers. Many, including Mr Harrison, spent time in crypto, once the most compute-intensive enterprise out there, which attracted financial engineers trying to turn processing power into a tradable asset. Mr Harrison was president of FTX US, the American arm of the crypto exchange founded by Sam Bankman Fried, resigning two months before the parent firm’s bankruptcy amid fraud in November 2022.
Luxor一度主要服务于比特币矿工。Anera的联合创始人之一维什瓦·奈克曾是加密货币创业者。（CoreWeave和其他一些新型云服务商也起步于加密货币挖矿业务。）随着机会增多，成熟的金融公司表现出更大兴趣。两大做市商DRW和Jump Trading均投资了Silicon Data，这家专注于算力研究的公司将与芝加哥商品交易所（CME）合作推出期货合约。
Luxor once primarily served bitcoin miners. One of Anera’s co-founders, Vishwa Naik, was a crypto entrepreneur. (CoreWeave and some other neoclouds also began life as crypto-mining businesses.) As opportunities grow, established financial companies are taking more of an interest. DRW and Jump Trading, two large market-makers, have both invested in Silicon Data, the compute-focused research firm with which CME hopes to launch its futures.
每家做市商也都组建了自己的团队来交易这一新商品。本月，Jump Trading收购了算力交易初创公司Ito Markets，其创始人将领导Jump进军这一新兴行业的工作。
Each market-maker has also assembled its own team to trade the new commodity. This month Jump Trading acquired Ito Markets, a compute-trading startup, whose founders will lead its efforts to break into the new industry.
Jump Trading投资主管大卫·奥尔森指出，有了合适的机构（如结算所）参与，市场将增长得更快，结算所通过确保交易双方履约来限制违约风险。“我们离那还很远，”他承认道，这呼应了业内许多人的看法——他们认为市场成熟可能需要数年甚至数十年。“但，”他补充说，“我们想参与这场讨论。”他的意愿并不令人意外。毕竟，面对依然稀缺的商品需求激增，且全球众多企业都暴露在其风险之下，奖赏极其丰厚。
David Olsen, Jump’s investment chief, points out that the market will grow more rapidly with the right institutions, such as clearing houses that limit the risk of blow-ups by ensuring that contracts are honoured by both sides of a trade. “We’re a long way from that,” he acknowledges, echoing many in this industry who think it may take years or even decades for the market to mature. “But”, he adds, “we want to be part of that discussion.”His desire is unsurprising. After all, with surging demand for what remains a scarce commodity, and with so many businesses around the world exposed to it, the prize is enormous.
Anera的奈克先生在进入加密货币领域前曾是印度大型企业集团信实工业的石油交易员，他将当下的算力市场时刻比作“七姐妹”末期。这七家西方巨头组成的卡特尔在20世纪20年代末至70年代间统治了全球石油贸易。它们的垄断地位因许多出口国的国有化以及几乎在所有地方兴起的公开石油市场而被打破。随着各国政府垂涎AI实验室，而像Anera这样的创业公司试图从零创建一个市场，你不必太费力就能看到算力将迎来类似的未来。
Mr Naik of Anera, who in his pre-crypto days was an oil trader at Reliance Industries, an Indian conglomerate, likens today’s moment in the compute market to the last days of the “seven sisters”, a cartel of Western giants that ruled the global petroleum trade between the late 1920s and the 1970s. Their chokehold was broken by nationalisation (in many exporting countries) and the rise of the open oil market (just about everywhere). As governments covetously eye the AI labs and ventures like Anera attempt to create a market from scratch, you do not need to squint too hard to see a similar future for compute.