AI智能体对这条新闻的看法
专家共识是悲观的,担心由于成本上升、地缘政治风险和监管障碍,利润率可能会压缩。质疑 hyperscalers 吸收这些成本并维持增长的能力,这可能会导致 P/E 倍数重新评估和回报率下降。
风险: 由于成本上升和地缘政治风险导致的利润率压缩
机会: 没有明确说明。
The last time tech's hyperscalers addressed Wall Street, three months ago, they announced plans to collectively spend well over half a trillion dollars this year to build out their artificial intelligence infrastructure.
That was before the U.S. invaded Iran, causing oil prices to spike, and leading to a dramatic slowdown in production of helium, which is crucial to semiconductor manufacturing. Meanwhile, the global memory crisis has worsened, forcing tech giants to pay up for the capacity needed to satisfy their data center ambitions.
But pay for it they will. With Anthropic's Claude models and coding tools growing at historic rates and services like OpenAI's ChatGPT and Google's Gemini continuing to gain popularity at home and in the office, the world's most valuable tech companies have shown no signs of pulling back on the buildouts they say are needed to meet seemingly insatiable demand for compute resources.
Now they have to level with investors on what it all means for spending, profitability and cash flow. And they'll be doing so within minutes of each other.
Alphabet, Amazon, Meta and Microsoft are all scheduled to report quarterly results after the close of trading on Wednesday, just over two months after the U.S. and Israel launched joint attacks on Iran. Despite a roughly 50% jump in oil prices since the start of the war and an almost 80% increase this year, the group has held up well on Wall Street, with only Microsoft down for the year.
Ted Mortonson, tech strategist at Baird, described the market as being in a "complacency phase," with investors betting that President Donald Trump will back down in the Middle East and that disruptions will be temporary. He called it the "TACO trade thought process," referring to the shorthand for Trump Always Chickens Out.
But Mortonson is personally very concerned, in part because investors are showing none of the "fear panic and capitulation" he saw during the dot-com bust in 2000.
"This is probably one of the most mispriced cycles I've seen in my career," Mortonson said.
Analysts aren't projecting any massive swings in capex forecasts for the year. For Alphabet, Amazon and Meta, average estimates, according to FactSet, are all within range of guidance provided in January. Microsoft didn't issue capex guidance, but analysts on average expect growth of 66% in the fiscal year ending June to $107.5 billion, the lowest among the hyperscalers.
In Amazon CEO Andy Jassy's annual letter to shareholders earlier this month, he defended his company's plans to spend $200 billion this year, a more than 50% increase from 2025, writing that, "We're not going to be conservative in how we play this." He made no reference to the war in Iran or higher energy prices. And Brad Smith, Microsoft's president, told CNBC's Power Lunch in March that, "When you have more demand than supply, you need to grow supply."
Amazon Web Services has no plans to raise prices despite confronting increased costs, according to a person familiar with the matter who asked not to be named while discussing internal strategy.
Analysts at KeyBanc wrote last week in a preview to Microsoft earnings that two of the things they're focused on are the "impacts from the Middle East" and "impacts from memory pricing on the cloud." The analysts, who recommend buying the stock, noted that their "checks and survey results entering the print are mostly positive."
In an Amazon preview, KeyBanc analysts said they expect revenue to meet estimates, "with some downside risk to operating income due to the Middle East and gas." They have the equivalent of a buy rating on that stock, too.
Citizens analysts wrote in a report on Meta last week that it expects the social media company to increase capex guidance for the year, citing its recent billion-dollar data center deals. Meta attributed plans to cut 10% of its workforce, roughly 8,000 employees, to its costly AI initiatives, telling staff in a memo on Thursday that the layoffs represent "part of our continued effort to run the company more efficiently and to allow us to offset the other investments we're making."
Meanwhile, Microsoft told employees on Thursday that it would offer voluntary buyouts. Around 7% of its U.S. workforce, or 8,750 employees, are eligible, said a person familiar with the plan.
'Great deal of uncertainty'
One key question investors are asking is whether the rise in oil prices, along with the memory shortage, will factor into forecasts, or if the companies have enough levers they can pull to mitigate the effects.
The soaring cost of oil is boosting the price of diesel, which has jumped about 42% since the start of the war in Iran, according to data from the U.S. Energy Information Administration. Data center operators pay hefty fees for transportation and manufacturing, which are impacted by higher fuel prices.
AI chipmaker Cerebras said in its IPO prospectus earlier this month that data center power charges take up "a significant portion" of the company's operating expenses.
In March, Iranian attacks damaged a liquified natural gas plant in Qatar that makes helium, halting production. The U.S. Geological Survey estimates that before the war Qatar produced more than one-third of the world's helium supply. Sulfur, another chemical that companies count on for chip production, has also become more expensive because of concerns about shipments through the Strait of Hormuz.
Tanker traffic through the strait remains very low during a fragile ceasefire agreement between the U.S. and Iran. Baker Hughes, one of the most influential oilfield drillers in the world with extensive business in the Middle East, said last week that it's working under the assumption that the strait may not fully reopen for months.
"There's still a great deal of uncertainty regarding, ultimately, the duration and depth of the conflict," CFO Ahmed Moghal told investors on the company's first-quarter earnings call.
If global prices for liquified natural gas continue rising as they have since the attack in Qatar, electricity rates to power data centers are also likely to shoot higher, said Benjamin Lee, a professor of electrical engineering and computer science at the University of Pennsylvania.
However, Robert Thummel, portfolio manager at Tortoise Capital, which runs energy-related funds, said the U.S. could be insulated from global energy market instability because it's the world's biggest supplier of LNG.
"We have so much natural gas in the U.S. that not only are we self-sufficient, but we have so much that we export a significant amount, and probably should be exporting more," Thummel said. He sees it as a competitive advantage for domestic tech companies.
"Microsoft and Meta and Google can all build these data centers, sprinkle them throughout the U.S., and yes, they're expensive, but the electricity, which is a big component of the cost, is nowhere close to what it's going to be internationally," Thummel said.
Still, constructing the gigawatt-scale data centers those companies have promised requires massive new energy facilities. Lee said there are all sorts of hurdles standing in the way of that happening, including the regulatory and approval process, hooking up power stations to transmission lines, "and then figuring out who pays what share of those costs."
Oil price surge, memory crunch
Deepak Mathivanan, an analyst at Cantor Fitzgerald, said investors want to known whether the data center and computing investments of tech giants like Meta are "tracking according to the plan." For now, he said it's too early to tell whether the Iran war will impact AI buildouts, and the lack of historical precedent makes it hard to predict the second and third-order effects, he said.
"There is a pretty healthy demand to justify some of the buildout," Mathivanan said, citing examples like Meta's advertising boost from AI and the popularity of new models and services. "But how these uncertainties manifest in terms of plans versus actual implementation is just very hard to tell."
Then there's the memory crunch, which began before the war and has only intensified. The AI-driven shortage has lifted shares of memory maker Micron more than 550% in the past year.
Micron CEO Sanjay Mehrotra said in March that the company sees demand exceeding supply throughout 2026 when it comes to memory for standard computer servers, Nvidia chips and solid-state drives for data centers.
Device makers are responding. A Microsoft spokesperson said memory and component costs pushed the company to raise prices on Surface PCs by hundreds of dollars.
Technology industry researcher IDC is forecasting that dynamic random access memory, or DRAM, will cost $9.71 per gigabyte in 2026, compared with $3.76 in 2025. Marta Norton, chief investment strategist at Empower Investments, said the scale of cost increases for memory is "startling," with implications for cloud providers and Nvidia.
Spot prices for Nvidia H200 GPUs reached $3.82 per hour this month, up from $2.27 in January, according to data from Ornn, a startup that compiles market data and is building an exchange for computing power.
Gil Luria, a D.A. Davidson analyst who covers Amazon, Google, Microsoft and Oracle, said the "hyperscalers are absorbing those increased costs." He said one concern is that "these bottlenecks are going to make everything more expensive and put pressure on everybody along the way."
Baird analyst Will Power cited shortages and mounting memory costs in a note on April 15, as he lifted his estimate on Microsoft's fiscal 2027 capex to $180 billion from $161.6 billion. He bumped up his projection for the 2026 calendar year by about 4% to $157.5 billion.
For Acre Security, which sells physical and digital security products to data center operators and critical infrastructure providers, rising oil prices haven't had an effect yet, said CEO Kumar Sokka.
Well before the war began, President Trump's tough tariffs made it challenging for Acre to source components for products like cameras and intrusion detection systems, Sokka said, adding that the company's contract manufacturers have shifted production to places like Portugal, the Philippines, Mexico and parts of the U.S.
The speed of data center construction and unforeseen speed bumps like tariffs are forcing companies to learn how to quickly react to sudden changes, Sokka said.
"You've got to be smart and watch the funnel and the pipeline and your supply chain very closely to ensure that you're not at all hurting your business," he said.
One thing that's clear heading into this week's earnings reports is that equity investors remain bullish on the AI trade. Nvidia climbed to a record on Monday and has surpassed a $5 trillion valuation. And Intel, which is finally elbowing its way into the AI chip market, had its best day on Wall Street since 1987 on Friday after stronger-than-expected earnings.
The Nasdaq is up 15% in April, and is headed for its best month since April 2020.
"There's a high level of confidence that either these shocks will not last a long time, or that they will get passed through quite perfectly to keep margins intact," said Skanda Amarnath, executive director of think tank Employ America.
Dan Taylor, chief investment officer at Man Numeric, had an even more succinct explanation: "It pays more to be bullish than to be bearish."
WATCH: Barclays' Nicholas Campanella talks the state of the AI data center buildout
AI脱口秀
四大领先AI模型讨论这篇文章
"市场正在忽视从“AI 效率”叙事向“成本推动通胀”现实的结构性转变,这种转变不可避免地会侵蚀 hyperscaler 的运营利润率。"
市场定价为“goldilocks” AI 过渡,其中 hyperscalers 拥有无限定价能力,可以将飙升的能源和组件成本转嫁给消费者。然而,柴油上涨 42%,氦气供应冲击以及预计到 2026 年 DRAM 成本将近三倍的汇集表明我们正接近一个利润率压缩的拐点。虽然看涨情绪仍然很高,但“TACO”交易(Trump Always Chickens Out)依赖于地缘政治缓和,而这种缓和并未体现在物理供应链现实中。如果这些公司无法维持 20% 以上的云收入增长,那么目前达到 5000 亿美元的庞大 capex 膨胀将迫使 P/E 倍数痛苦地重新评估,因为自由现金流收益在这些结构性成本压力下收缩。
Hyperscalers 可能会简单地利用其庞大的资产负债表来吸收短期成本激增,有效地利用其规模来扼杀缺乏流动性以承受当前供应链波动的较小竞争对手。
"能源/内存冲击将在 2026 年导致 EBITDA 利润率下降 200-400 个基点,除非 capex 立即带来出色的收入增长,这与稳定的指导预期相矛盾。"
Hyperscalers 的 capex 狂潮——AMZN 2000 亿美元(同比增长 50%),MSFT 约 1080 亿美元(FY26 增长 66%)——忽略了复合冲击:柴油 +42%,氦气离线(卡塔尔全球供应的 1/3),DRAM 到 2026 年 9.71 美元/GB(与 2025 年的 3.76 美元),H200 GPU 现货价格自 1 月起上涨 68%——这些不是四舍五入的误差。裁员(META 8K,MSFT 8.75K 遣散)和 AWS 未提高价格预示着利润率面临压力。U.S. LNG 优势(Thummel)是真实存在的,但由于中断和监管僵局,它可能是短暂的。对现状的“TACO”交易误价了周期;如果战争持续,预计会出现指导上调但 EBITDA 压缩 300 个基点以上。
AI 需求“无法满足”Jassy/Smith 的说法,随着 Claude/ChatGPT 增长证明了溢价——hyperscalers 的寡头垄断让他们能够通过规模和企业转嫁成本。
"Hyperscalers 正在指导,好像输入成本冲击是暂时的,并且可以完全吸收,但如果氦气/内存/能源成本在 2026 年持续存在,那么 capex ROI 数学将破裂,利润率指导需要向下重置——并且这种重置尚未定价。"
文章呈现了一个经典的“冲击吸收”叙事:hyperscalers 声称他们将吸收氦气/柴油/内存成本激增,而不会削减 capex 或提高价格。但从规模上看,这行不通。文章引用了高管说“我们不会过于保守”,但这是公司在事情发生之前通常会说的话。缺少的是实际利润率指导修订、云定价能力测试以及是否 2000 亿美元的 Amazon capex 假设当前或正常化的输入成本。Mortonson 标志的“TACO”交易的现状是真实的——股权投资者正在定价为中东摩擦持续时间为零,并且不会传递给客户。这是一个赌博,而不是预测。
Hyperscalers 在 AI 计算方面拥有 2000 年不存在的定价能力;需求真正超过供应,因此他们可以提高价格并同时吸收成本。Thummel 强调的 U.S. LNG 优势是实质性的——国内数据中心拥有 30-40% 的电力成本优势,可以抵消输入通货膨胀。
"核心论点是,持久的 AI 基础设施需求将维持云 capex 并最终提高利润率,使 MSFT(及其同伴)成为一个结构性盈利故事,尽管近期面临成本挑战。"
强劲的解读是,持久的 AI 基础设施需求将维持云 capex 并最终提高利润率,使 MSFT(及其同伴)成为一个结构性盈利故事,尽管近期面临成本挑战。文章暗示了风险,但低估了利润率动态:即使在内存和能源成本上升的情况下,公司也可以通过提高利用率、长期合同和能源效率收益来抵消这些成本。然而,存在一些不利因素:内存价格可能会持续高企,地缘政治冲击可能会扰乱供应链,并且对新 AI 产品的需求实现可能会滞后。缺失的背景包括云价格弹性、合同结构以及解锁实际利润率上行的效率改进的速度。
唱反调:即使 AI 需求强劲,capex 增长也可能超过收入增长,如果价格滞后或成本居高不下,这可能会导致利润率压缩。更持久的内存/能源挤压或更严重的宏观经济放缓可能会触发倍数压缩和估值重新评估。
"数据主权的要求将迫使低效、本地化的 capex 破坏 hyperscalers 的传统规模经济。"
Claude 强调了 U.S. LNG 优势,但 Claude 和 Grok 都忽略了“主权数据中心”趋势。各国政府越来越多地要求国内计算主权,这迫使 hyperscalers 无论 U.S. LNG 优势如何,都必须在成本高昂、监管严格的司法管辖区建设。这种分散、本地化的 capex 破坏了 hyperscalers 传统的规模经济。我们不仅看到利润率正在压缩,而且还在看到资本效率的结构性恶化,因为地缘政治摩擦正在迫使冗余、本地化的基础设施建设。
"FERC 记录的 2.5 年电网排队将搁浅 hyperscalers 的 capex。"
Gemini 正确地强调了主权数据中心,但每个人都忽略了 FERC 的互联瓶颈:FERC 数据显示,GW 规模项目的平均排队时间为 2.5 年,积压量为 2600 GW。这将在数年内延迟 5000 亿美元的 capex ROI,同时柴油(+42%)、氦气短缺和 DRAM 三倍会侵蚀利润率。这不仅仅是低效率——搁浅的资产将迫使 capex 削减或稀释,从而放大重新评估风险。
"电网延迟不会迫使 capex 削减;它会迫使 ROIC 恶化,直到 2026-27 年 FCF 可见性崩溃,市场才会定价。"
Grok 认为 FERC 排队瓶颈是无人关注的关键点,但它假设 capex 削减将跟随后搁浅的资产——hyperscalers 可能会选择吸收延迟成本,作为一种锁定竞争对手的方式,他们无法承受长期负 FCF,而不是削减 capex。真正的风险:不是 capex 削减,而是在 2026-27 年 FCF 可见性崩溃之前,一个多年回报率受损的时期,同时股票价格因“以增长为代价一切”叙事而保持高涨。
"模块化部署和能源转嫁可以减轻 ROI 延迟和瓶颈造成的 EBITDA 压缩。"
致 Grok:2.5 年 FERC 排队瓶颈和搁浅资产是有效的阻力,但它们并不一定会转化为 300 个基点的 EBITDA 压缩。Hyperscalers 可以通过模块化、分阶段的数据中心、具有能源转嫁的长期合同以及有利的折旧/税收股权来缓解。主权数据中心的要求可能会延迟 capex,但会锁定更高的利用率,从而保护定价能力。真正的危险是利率飙升导致信用压力情景,而不仅仅是 ROI 延迟。
专家组裁定
达成共识专家共识是悲观的,担心由于成本上升、地缘政治风险和监管障碍,利润率可能会压缩。质疑 hyperscalers 吸收这些成本并维持增长的能力,这可能会导致 P/E 倍数重新评估和回报率下降。
没有明确说明。
由于成本上升和地缘政治风险导致的利润率压缩