在过去四年里，作为美国能源经济心脏的达拉斯，已成为美国数据中心经济的心脏。这得益于其廉价的土地、友好的监管姿态、流动的劳动力市场、优越的交通基础设施以及充足的电网接入。大都会地区约有125座设施正在运营，另有19座在建、53座规划中。这场建设热潮正推高蓝领工人的工资，并帮助将当地失业率维持在4.1%，接近历史低位。
Over the past four years, Dallas, the beating heart of the U.S. energy economy, has become the beating heart of the U.S. data-center economy. That’s thanks to its cheap land, friendly regulatory posture, liquid labor market, excellent transportation infrastructure, and ample grid connectivity. Roughly 125 facilities are up and running in the metro region, and an additional 19 are under construction and 53 are planned. The building boom is boosting wages for blue-collar workers and helping hold the local unemployment rate at 4.1 percent, close to its historic low.
然而，达拉斯及其郊区同时也处于一场建设低迷之中。零售建设势头强劲。但尽管租金高企、人口增长，新建住宅却寥寥无几：开工房屋数量同比下降12%，在建多户住宅单元数量在过去三年暴跌43%。办公楼建设同样低迷。
Yet Dallas and its suburbs are also in the midst of a building bust. Retail construction is strong. But despite high rents and a growing population, few new homes are being built: The number of houses builders broke ground on dropped 12 percent year-over-year, and the number of multifamily units in the works plummeted 43 percent over the past three years. Office construction is also depressed.
科技繁荣是抵消了这些其他领域的低迷，还是导致了它们？这是达拉斯乃至全美面临的重大问题。
Is the tech boom offsetting these other downturns or causing them? That’s the big question facing Dallas and the rest of the United States.
投资者对人工智能的热情，自铁路时代以来从未有过。若非他们将洪流般的资金导向那寥寥几家打磨算法、建设数据中心的公司，美国的就业水平和经济产出可能会更低。但反过来说，若非他们将洪流般的资金导向那寥寥几家打磨算法、建设数据中心的公司，这些资金本可能流向其他所有人和所有事物。那么，人工智能究竟是为美国经济注油，还是在抽走它的油水？正如经济学家所言，它是“挤出”了非科技企业的投资，还是在一个原本疲软的扩张期中提振了投资？
Investors are more pumped about AI than they have been about any technology since the railroad. If they weren’t directing a torrent of money to the handful of firms finessing algorithms and constructing data centers, the country’s employment level and economic output might be lower. Then again, if they weren’t directing a torrent of money to the handful of firms finessing algorithms and constructing data centers, they might be directing it to everyone and everything else. So is AI fueling the American economy or siphoning gas from it? Is it “crowding out” investment in non-tech enterprises, as economists put it, or boosting investment in an otherwise lackluster expansion?
毫无疑问，AI正在推高GDP——或者更确切地说，是AI正在推高GDP。2020年，美国企业在耐用资产投资上（如软件、仓库和卡车）花费了3.6万亿美元（按通胀调整后）。去年，这一支出达到4.4万亿美元。甲骨文、微软、亚马逊、Meta和Alphabet仅今年就在资本支出上投入8000亿美元，这一数字在过去十年间增长了八倍。
Certainly, AI is pushing up GDP—or, to be precise, the AI is pushing up GDP. In 2020, American businesses spent $3.6 trillion, adjusting for inflation, on investments in durable assets, such as software, warehouses, and trucks. Last year, they spent $4.4 trillion. Oracle, Microsoft, Amazon, Meta, and Alphabet are shelling out $800 billion on capital expenditures just this year, a sum that has grown eightfold over the decade.
这是一场硅谷版的“硬帽刺激计划”，规模甚至超过华盛顿为应对大衰退而通过的刺激方案。它正在改写国民账户。今年第二季度，家庭消费以3.8%的年化速度增长。除房地产外的私人投资以12.7%的速度攀升。设备支出以21%的速率激增。仅计算机采购就为GDP贡献了0.7个百分点，“信息处理设备”更贡献了高达0.9个百分点。
It’s a Silicon Valley hard-hat stimulus, bigger than the one Washington passed to counter the Great Recession. And it is transforming national accounts. In this year’s second quarter, household consumption rose at a 3.8 percent annualized pace. Private investment in everything but real estate climbed at a pace of 12.7 percent. Spending on equipment went up at a 21 percent rate. Purchases of computers alone added 0.7 percentage points to GDP, and “information processing equipment” a mammoth 0.9 percentage points.
这种私营部门刺激与纽约、伦敦和上海的关系，不亚于与库比蒂诺和旧金山的关系。早期的AI繁荣主要由极其富有的风险投资人和利润丰厚的科技公司资助，许多投资者投入了现金。但很快，AI企业为完善模型需要更多算力。它们需要建设数据中心和芯片工厂。它们需要的流动性比湾区能提供的多出数千亿美元。全球投资者看中它们承诺的耀眼回报，纷纷入场。随着贷款需求上升，贷款成本攀升。挤出效应反而引来了挤入效应。
This private-sector stimulus has as much to do with New York, London, and Shanghai as it does with Cupertino and San Francisco. The early AI boom was mostly financed by wildly wealthy venture capitalists and wildly profitable tech companies, many of the investors staking cash. But soon, AI firms needed more computing power to perfect their models. They needed to build data centers and chip factories. They needed hundreds of billions of dollars more liquidity than the Bay Area could provide. The world’s investors took a look at the sparkling returns they were promising and signed on. The cost of loans climbed as demand for loans climbed. Crowd out crowded in.
这种影响在建筑业最为明显。据哈佛大学经济学家杰森·弗尔曼的分析，与人工智能相关的企业目前正在建设543个数据中心，另有2513个处于筹备阶段。过去一年，数据中心建设支出增长了65%。与此同时，用于住房和零售店建设的资金下降了10%。教育设施建设支出下降13%；办公楼下降14%；酒店、汽车旅馆和度假村下降15%；制造工厂下降24%。
The effect is easiest to see in the construction sector. AI-linked firms are currently building 543 data centers, and 2,513 more are in the pipeline. Spending on the build-out has increased 65 percent over the past year, according to an analysis by the Harvard economist Jason Furman. At the same time, the amount devoted to building housing and retail outlets has dropped 10 percent. Spending on the construction of education facilities is down 13 percent; offices, 14 percent; hotels, motels, and resorts, 15 percent; and manufacturing plants, 24 percent.
科技企业不仅让非科技企业为建设项目融资的成本更高。它们还在推高劳动力、土地和材料的成本。它们刺激房东和机构投资者购买国债或股票，而不再费心投资公寓楼和养老院。它们也迫使家庭为住房贷款支付更多费用。总的来说，高收益率正在“把所有资金都从抵押贷款市场吸走”，Redfin首席经济学家达里尔·费尔韦瑟告诉我。
Tech enterprises are not merely making it more expensive for non-tech enterprises to finance construction projects. They’re bidding up the cost of labor, land, and materials. They’re spurring landlords and institutional investors to buy Treasuries or stocks instead of bothering with apartment buildings and retirement homes. They’re forcing families to pay more for home loans too. As a general point, high yields are “sucking all the money out of the mortgage market,” Daryl Fairweather, the chief economist at Redfin, told me.
试图建设基础设施的政府也遭遇了同样的借贷成本上升。想要购买新车或翻新房产的家庭也是如此。想要创办小企业或报名培训项目的年轻人也是如此。以一种迂回的方式，Claude和ChatGPT可能也在推高信用卡逾期率。
Governments trying to build infrastructure are getting hit by the same increase in borrowing costs. So are families looking to buy a new car or renovate a property. So are young people looking to launch a small business or sign up for a training program. In a roundabout way, Claude and ChatGPT might be increasing credit-card delinquencies as well.
诚然，在建筑和房地产这两个对利率和间接费用高度敏感的行业之外，这种影响可能并不那么大。（贸易战、打击移民以及与伊朗的战争也在重创这两个行业。）密歇根大学经济学家贾斯汀·沃尔弗斯指出，“2007年至2022年是一个不同寻常的时期”，当时借贷成本接近零。如今，利率已回到1990年代的水平，且仍低于1980年代。
Granted, the effect might not be that big outside of the construction and real-estate sectors, which are highly sensitive to interest rates and overhead expenses. (The trade war, immigration crackdown, and war with Iran are pummeling them too.) Justin Wolfers, an economist at the University of Michigan, noted that “2007 through 2022 was an unusual period” in which borrowing costs were close to zero. Today, rates have returned to where they were in the 1990s, and they remain lower than they were in the 1980s. They are “what I would call normal,” he said.
他说，这是“我所谓的正常水平”。此外，较高的利率迫使投资者在下注前三思，首先挤掉了最投机、最荒唐和最骗人的项目。斯坦福经济政策研究所经济学家瑞安·卡明斯告诉我，加密货币挖矿设施“显然正被挤出”，他称这是“件好事”。但话说回来，住房和公寓楼是好事，工厂和酒店也是好事。由人工智能驱动的利率微小变化，可能导致数十万美国家庭在未来十年继续做租客。可能导致大量商业网点无法建成。可能导致数百家非科技初创企业仅停留在构想阶段。
Moreover, higher interest rates force investors to think before they stake, squeezing out the frothiest, goofiest, and scammiest projects first. Cryptocurrency mining facilities are “clearly being crowded out,” Ryan Cummings, an economist at the Stanford Institute for Economic Policy Research, told me. He called that “a good thing.”Then again, houses and apartment buildings are a good thing, as are factories and hotels. Small, AI-driven changes in interest rates might lead to hundreds of thousands of American families remaining renters for the next decade. They might lead to scores of commercial outlets not getting built. They might lead to hundreds of non-tech start-ups remaining nothing more than an idea.
这些数据中心除了对人工智能公司有用之外，真的会是件好事吗？或许不会。通常，当城市和城镇吸引公司在当地建设设施时，它们会变得更加繁荣，信用也更好。“这是一个二加二等于五的论点，”宾夕法尼亚州立大学金融学教授杰斯·科纳吉亚告诉我。发电厂、学校、体育场和工厂会雇佣工人，为当地企业创造收入，增加税收，并促使更多公司聚集在一起。
Will all of these data centers be a good thing, aside from their utility to AI firms? Perhaps not. Normally, cities and towns become more prosperous and credit-worthy when they lure companies to build facilities there. “It’s a two-plus-two-equals-five argument,” Jess Cornaggia, a finance professor at Pennsylvania State University, told me. Power plants, schools, stadiums, and factories hire workers, generate revenue for local businesses, increase tax receipts, and spur additional firms to co-locate.
但数据中心最终可能是一个二加二等于四，甚至二加二等于三的命题。投资者要求在拥有大量数据中心的社区，为水利和学校项目债券提供更高的收益率，这意味着他们预计这些社区未来将面临更多资源压力，并征收更少的房地产税收入。毕竟，一旦数据中心开始运行，它们会消耗大量电力和水资源，却几乎不创造就业岗位。
But data centers might end up being a two-plus-two-equals-four, or even two-plus-two-equals-three, proposition. Investors are demanding larger yields on bonds for water and school projects in communities with a lot of data centers, meaning they expect those communities to experience more resource strains and collect less property-tax revenue in the future. Once data centers are up and running, they use a considerable amount of power and water while generating close to no jobs, after all.
这些权衡在关于人工智能对福祉和财富影响的炒作盛行的辩论中大多缺席，但它们很重要。如果白领公司变得极其高效和富有创造力，放弃住宅、公园、健康诊所和杂货店可能是值得的。如果我们能攻克癌症，放弃非科技初创公司可能也是值得的。但国家在人工智能上投入的资源越多，以牺牲其他创新为代价，其未来就越依赖于这一赌注能否成功。
These trade-offs have been mostly absent from the hyped-up debate over AI’s effect on welfare and wealth, but they’re important. Giving up homes, parks, health clinics, and grocery stores might be worth it if white-collar firms become radically more productive and creative. Giving up non-tech start-ups might be worth it if we get a cure for cancer. But the more resources the country devotes to AI at the expense of other innovations, the more its future depends on that bet paying off.