For two years, Jim Covello has been asking the same uncomfortable question: when does Wall Street’s AI bet actually pay off?
The head of global equity research at Goldman Sachs raised the alarm in a widely read 2024 report, questioning whether the torrent of capital flooding into artificial intelligence would ever generate returns commensurate with the spending. At the time, he estimated it would take 18 months to 2 years. Two years later, speaking on Goldman’s Exchanges podcast, he noted that he’s been having this debate for close to four years now — and his answer has only grown more pointed.
“At some point, you’ve got to make money,” Covello said. “You make investments in a business so that you can generate returns and make money. And we’ve gotten further away from that over the last couple years instead of closer to it.”
Speaking with Allison Nathan, Goldman’s senior macro strategist, and George Lee, co-head of the Goldman Sachs Global Institute — who has long taken a more optimistic view of AI’s potential — Covello laid out three reasons why that patience may be running thin and asked another question that runs counter to the market narrative.
“If we’re having the same debate two years from now and we’re still saying, ‘Well, it’s early,’ then we might have a challenge,” Covello said, “because at some point, when does the short-term become the long-term?”
Anthropic and OpenAI weren’t mentioned specifically on this podcast, but they are both nearing mega-IPOs, they are both valued at close to $1 trillion, and neither is profitable.
The hill keeps getting steeper
Covello acknowledged being wrong on some counts. Consumer adoption of AI has been, in his words, “magnificent” — far exceeding his expectations. And the technology itself has advanced rapidly. But on the question that matters most to investors, his conviction has hardened rather than softened.
“In a lot of ways, companies are losing more money today implementing this technology than they were two years ago,” he said. “The hill that has to get climbed is even steeper today than it was before, because we’ve spent more money.”
Goldman Sachs Global Institute, which takes a more optimistic long-term view of AI, nonetheless agrees the math is daunting. Co-head George Lee estimated that $7 trillion to $8 trillion will ultimately be spent on AI infrastructure — and that simply disrupting existing profit pools won’t generate sufficient payback. Net new economic activity, Lee argued, is the only way the numbers eventually work.
The critical variable for both men is the same: enterprise ROI. And that, Covello said, remains the central unanswered question. “I really think it all boils down to one thing: do the enterprises make or save money implementing AI? If they do, this technology is going to fulfill its promise.” But if, two years from now, people are still saying it’s “early” in the adoption wave, “then we might have a challenge.”
So far, the evidence is thin. Widely cited MIT research found 95% of organizations reporting zero return on AI pilots. A 2025 EY survey found 99% of companies reported financial losses tied to AI-related risks, averaging $4.4 million per company.
The data tracks with what Fortune heard directly from executives at its COO Summit earlier this week. Cognizant, whose research team presented new findings at the event, reported that 93% of jobs are already being disrupted by AI — six years ahead of their own 2023 projections — and yet the productivity gains that were supposed to follow haven’t materialized. Their researchers called it an “activation gap.” That dynamic surfaced in the executive debates, too: Francine Katsoudas, EVP and Chief People Officer at Cisco, noted that on teams using AI most intensively, trust within those teams began to drop about 9 months in — a warning sign that the human-organizational dimension of AI deployment remains deeply unsolved. “We just have to invest so much more,” she said.
The capex paradox
Part of what makes Covello’s frustration notable is the sheer scale of spending without proof of return. Hyperscalers — Amazon, Microsoft, Google, Meta — have not only maintained their AI capital expenditure despite stock underperformance, but they’ve also increased it. Covello predicted two years ago that sustained underperformance would trigger spending discipline. It didn’t.
“There’s a tremendous amount of FOMO at every level of the supply chain,” he said, describing a dynamic where every company, from enterprise to model layer to hyperscaler, fears being left behind if a competitor cracks the economic code first. (The FOMO theme has been on Covello’s mind in recent months.)
At Fortune‘s Most Powerful Women Summit last October, Brilliant Earth chief brand officer Pam Catlett put it plainly: “For me, there’s the pervasive sense of FOMO that’s happening — fear of missing out. And the first word, fear, is not a good state to be in when you’re thinking about how to better serve your customer.”
The result, Covello said, is an unusual inversion in the supply chain. Semiconductor companies — led by Nvidia — are capturing nearly all the economic value being created, while the companies above them are bleeding cash. Covello said this is historically unprecedented — and he covered semiconductor stocks for 16 years. “In every cycle, the semiconductor stocks thrive when their customers thrive. Here in this cycle, the semiconductor companies are thriving at the economic expense of everybody above them in the chain.”
He now favors hyperscaler stocks over semiconductors — a notable reversal — arguing that in two of three likely scenarios, hyperscalers win. Only a pure status quo, where semis alone keep profiting indefinitely, keeps the chip trade alive. “That can’t go on forever,” he said.
Workers aren’t feeling it
A second pressure front is building inside companies themselves. Third-party surveys consistently document a widening gap between C-suite enthusiasm for AI and what workers on the ground are actually experiencing. Covello pointed to the pattern bluntly: “The line workers aren’t getting as much benefit from it as the C suite expected.”
Part of the explanation, he said, is a data readiness problem that rarely makes headlines. “There are agents today that are terrific. There are models today that are terrific. But in many cases, the data isn’t ready to be agented yet. So we’re putting agents on top of data that isn’t ready to be agented. And that’s creating another economic challenge for companies.”
Large incumbent companies also face a structural drag that AI-native startups don’t, Goldman’s Lee noted — legacy systems, entrenched workflows, and resistance to change create a “drag coefficient” that slows the realization of productivity gains at scale. The productivity leap is real, he argued, but primarily visible in companies built from scratch for this technology. Getting there for everyone else will take longer.
These are problems that Yale’s Jeffrey Sonnenfeld examined in depth in Fortune last month. In a piece titled “Why your data infrastructure — not your AI model — will determine whether agentic AI scales,” the Lester Crown Professor in the Practice of Management argued that the bottleneck in enterprise AI deployment isn’t model quality but the organizational plumbing beneath it — siloed systems, inconsistent data governance and infrastructure built for a pre-AI world that simply wasn’t designed to support autonomous agents making real-time decisions. Part of a series on agentic AI’s enterprise implications, Sonnenfeld and his colleague Stephen Henriques found that data infrastructure readiness varies dramatically across sectors, with most large incumbents badly underestimating what it would take to make their data “agentable” — a word Covello, independently, used almost verbatim.
Politics enters the equation
Perhaps the most underappreciated pressure is the one building outside corporate boardrooms entirely. Lee raised it unprompted: a “somewhat sudden turn towards populist resentment of AI” that he described as “a deeply unpopular technology — interestingly, almost uniquely in the US versus other parts of the world.”
The signs are obvious. Mentions of AI are booed at commencement addresses by Gen Z sensing their futures being erased. Data center deployment has triggered threats and protests in local communities facing rising electricity costs.
President Trump convened seven major hyperscalers at the White House in March to sign a voluntary “Ratepayer Protection Pledge” — a political signal of just how quickly the issue has escalated. Data centers now consume over 4% of U.S. electricity, with some government projections suggesting that could triple to 12% by 2028.
Covello connected the politics directly to the economics. “I would put all of it in the broader economics bucket,” he said. “How much of it is going to flow back to the individual? Can we make a case that individuals are benefiting economically from using the technology?” Clearly, the implication is that the case is not being made.
The market, for now, is still granting a long leash — buoyed by a bull market that Covello himself acknowledged is partly caused by AI optimism, creating a circular dynamic that’s difficult to unwind. But the patience is not unconditional, and three clocks — economic, operational, and political — are now running simultaneously.
“That doesn’t mean it’s never going to happen,” Covello said of AI’s eventual payoff. “It just means the stakes are higher.”
The post ‘At some point you’ve got to make money’: Goldman’s top AI skeptic warns the clock is running out ahead of OpenAI and Anthropic IPOs appeared first on Fortune.
