There's a pretty big disconnect between demand and compute availability
In Hong Kong this week, a Goldman Sachs executive offered a measured defense of one of the largest capital commitments in the history of technology: the roughly $700 billion American firms have poured into AI infrastructure this year. Eric Sheridan's argument rests not on faith in future potential but on a present-tense imbalance — demand for computational power already outstrips supply, and agentic AI systems capable of autonomous, economically productive work are beginning to close the gap between promise and utility. The deeper question the moment poses is whether technological abundance, once built, can outrun the forces — cheaper rivals, compressed margins, unproven returns — that always follow in its wake.
- A $700 billion infrastructure bet is under fire as investors question whether AI spending will ever generate returns proportional to its scale.
- Chinese open-source models are arriving cheaper and faster, threatening to erode the profit margins of American firms that built expensive proprietary systems.
- Goldman Sachs argues the shift to agentic AI — systems that act autonomously, not just respond — marks the moment speculation becomes measurable economic utility.
- Compute demand is expected to outpace supply well into late 2027, effectively locking in continued heavy infrastructure spending regardless of the debate.
- The unresolved tension: if open-source alternatives prove sufficient for most use cases, the West's most expensive infrastructure build may yield less than it cost.
Eric Sheridan arrived in Hong Kong this week with a case to make on behalf of an enormous wager. Speaking at Goldman Sachs' Asia Communacopia + Technology Conference, the firm's executive framed the current moment as an inflection point — the arrival of agentic AI, systems capable of autonomously performing complex tasks, that he believes will finally justify the staggering sums American technology companies have committed to AI infrastructure.
Those sums are difficult to absorb: US tech firms are on pace to spend $700 billion this year alone on data centers, semiconductors, and the physical backbone of artificial intelligence. The spending has drawn sustained skepticism from investors, and that skepticism has sharpened as Chinese companies release cheaper, open-source AI models that threaten to undercut the proprietary advantages American firms paid so dearly to build.
Sheridan's rebuttal centers on supply and demand. Computational capacity, he explained, cannot keep pace with what companies now want to use — and that imbalance will persist well into the second half of 2027. In his framing, this shortage is itself the argument: if demand reliably exceeds supply, the infrastructure will keep being built, and those who built the most will hold the advantage longest.
What the argument cannot yet resolve is whether agentic AI will deliver productivity gains visible enough to satisfy investors — or whether cheaper open-source alternatives will erode the returns before the infrastructure pays for itself. For now, Sheridan's position is that the race belongs to whoever commands the most compute, and that lead, however contested, still belongs to those who spent the most to secure it.
Eric Sheridan stood in Hong Kong this week with a straightforward message for skeptics: the massive bet American technology companies have placed on artificial intelligence infrastructure is about to pay off. The Goldman Sachs executive, speaking at the firm's Asia Communacopia + Technology Conference on Monday, framed the moment as a turning point—the arrival of agentic AI tools that actually produce economic value, not just theoretical capability.
The numbers behind this wager are staggering. US tech companies are on pace to spend $700 billion this year alone on the foundational machinery of AI: data centers, semiconductors, the physical and computational backbone that makes everything else possible. That figure has drawn sustained criticism from investors and analysts who wonder whether such historic spending will ever generate proportional returns. The skepticism is not abstract. It arrives with real competition: Chinese companies have begun releasing cheaper, open-source AI models that threaten to compress the profit margins of American firms that spent lavishly to build proprietary systems.
Sheridan's argument cuts through this doubt with a specific claim about supply and demand. Right now, he explained to the South China Morning Post, there is a significant gap between how much computational power companies want to use and how much is actually available. This imbalance, he suggested, will not resolve itself quickly. The shortage of compute capacity will persist well into the second half of 2027—a timeline that effectively guarantees continued heavy spending on infrastructure for at least another year and a half.
The logic is straightforward: if demand for computing power exceeds what exists, companies will keep building data centers and buying chips. The arrival of agentic AI—systems that can autonomously perform complex tasks rather than simply respond to prompts—has shifted the conversation from whether AI will be useful to whether there is enough hardware to meet the demand. This is the inflection point Sheridan identified: the moment when the technology stops being a speculative bet and starts being a practical tool that organizations actually want to deploy at scale.
What remains unresolved is whether this productivity will translate into the financial returns that justify the spending. The open-source competition from China complicates the picture. If cheaper alternatives become viable for most use cases, American companies may have built more capacity than they can profitably fill. But Sheridan's framing suggests that the immediate future belongs to whoever has the most compute available—and that advantage, for now, belongs to those who spent the most to build it.
Citas Notables
There's a pretty big disconnect between the demand and the availability of compute. We don't think that imbalance closes until well into the second half of 2027.— Eric Sheridan, Goldman Sachs
La Conversación del Hearth Otra perspectiva de la historia
When Sheridan talks about an inflection point, what exactly is he saying has changed?
He's saying that agentic AI—systems that can actually do work autonomously—has moved from theoretical to real. Companies now want this technology, not just as a curiosity but as something they'll deploy at scale. That creates genuine demand for computing power.
But doesn't that demand eventually get satisfied? Once you build enough data centers, the shortage ends.
Yes, but he's arguing that won't happen until late 2027. That's a long runway of sustained spending. By then, the companies that built the most capacity will have a structural advantage.
What about the Chinese open-source models undercutting prices?
That's the real tension. If those models become good enough for most tasks, American companies may have overbuilt. But right now, the scarcity of compute is so acute that whoever has capacity available wins. The open-source threat matters more once supply catches up.
So this is really about timing?
Exactly. Sheridan is saying the companies that spent the most are protected by scarcity—at least for the next 18 months. After that, the economics get much harder.
And if agentic AI doesn't deliver the productivity gains he's assuming?
Then $700 billion in spending looks like a very expensive mistake. But he's betting that the demand he's seeing now is real, not hype.