The Iran war has laid bare a paradox: Gulf money is helping underwrite America’s effort to win the artificial intelligence race, and now the US has started a conflict that could destabilize those investments. Some estimates have projected $2 trillion in long-term pledges from Middle Eastern nations to the AI boom, money that now looks precarious. At the same time, surging energy costs threaten to make data centers far more expensive to run. But the aftershocks of the conflict appear less likely to kill the AI boom entirely than cleave the market in two, leaving so-called hyperscalers like Alphabet Inc., Amazon.com Inc. and Microsoft Corp. most exposed to the shifting financial landscape while upstart AI labs such as OpenAI and Anthropic PBC are more insulated.

Investors have long treated the AI bonanza as a single, monolithic story, but in reality it has two distinct elements — a phenomenally expensive infrastructure business, and a cheaper software play. Among the architects of the latter component, Anthropic has been chugging along rather well lately with annualized revenue more than doubling in the last three months to $19 billion, while OpenAI’s is at around $25 billion. Consumers, business clients across finance and life sciences and governments are all paying for subscriptions and access; unlike previous hype cycles around the metaverse and crypto, that momentum looks sustainable.

For all the worries about OpenAI’s high cash-burn rate, the AI labs also benefit from sticky enterprise contracts. Clients are unlikely to cancel these because of geopolitical uncertainty; instead they’re likely to maintain them in the hope of making their organizations efficient enough to ride whatever choppy economic waves may be incoming.

The AI software makers need data centers to run their businesses, but they’re not directly exposed to rising energy costs in the way the owners of those server farms are. To make money, OpenAI and Anthropic need to run their existing AI models to answer queries from their paying customers, a process known as inference. But training new frontier models is much more energy intensive, requiring the continuous use of thousands of AI chips (graphics processing units or GPUs made by Nvidia Corp.) for weeks or months on end.

Of course, the day-to-day costs of inference add up over time, especially for a company like OpenAI, which claims 900 million weekly users. But the energy load is much lower, more distributed and easier to manage than training the next generation of models, something the labs can afford to delay while they focus on urging businesses to plug existing tech into their workflows.

Hyperscalers like Amazon, Google, Microsoft, Meta Platforms Inc. and Oracle Corp. are more at risk given how much their $1.15 trillion buildout relies on cheap, reliable energy, especially natural gas. It’s the dominant single energy source for US data centers, providing about 40% of their power, according to the International Energy Agency — a problem when the Iran war is driving up prices.

The chip supply chain is similarly exposed. Taiwan Semiconductor Manufacturing Co. (TSMC) makes nearly all the high-end chips designed by Nvidia, but Taiwan also depends on the Middle East for about a third of its fuel, and the island nation gets most of its helium from Qatar. The gas is critical in semiconductor manufacturing thanks to its unique ability to cool and protect silicon wafers during production. Helium production at QatarEnergy’s Ras Laffan Industrial City was crimped last week following an Iranian drone attack; the broader implication could be a months-long wait for chip output to fully recover.

That leaves Nvidia perhaps the most exposed of all. The world’s most valuable public company, with a market cap exceeding $4 trillion, derives most of its revenue from selling chips to hyperscalers. Anything that slows down the buildout of vast new server farms will hurt its order book.

While Alphabet and Amazon at least have recurring cloud subscriptions to act as a financial cushion, Nvidia doesn’t have any such revenue stream. It just sells chips, which face the double whammy of being harder to manufacture in Taiwan in addition to a question mark over recent mega-deals with the Middle East. In November, the U.S. government approved Nvidia’s sale of 70,000 of its most advanced chips to the United Arab Emirates and Saudi Arabia, a deal that now looks more uncertain. (Nvidia declined to comment.)

Energy and cash from the Gulf have helped fuel the AI boom. However strong the revenue growth is for its applications, the outlook for the underlying infrastructure looks ever more fragile the longer the war carries on.