Goldman’s Jim Covello warns AI IPO hype won’t outrun ROI reality

Jim Covello has been poking the same sore spot for two years at Goldman Sachs: when does AI spending turn into money, not headlines? On Goldman's Exchanges podcast, the head of global equity research sharpened his argument to the point of urgency, saying, "At some point, you've got to make money," and that Wall Street has moved farther away from the payoff instead of closer.

He also asked the question that makes executives sweat: "When does the short-term become the long-term?" Covello’s timing matters because the market narrative is already sprinting ahead. Fortune notes that Anthropic and OpenAI were not mentioned on the podcast, but both are nearing mega-IPOs, both are valued at close to $1 trillion, and neither is profitable. In other words, the question is not academic. If the “it’s early” refrain stretches too long, the IPO math gets shakier for everyone watching.

Covello is not saying AI is a dud. He acknowledges two key positives: consumer adoption of AI has been "magnificent," exceeding expectations, and the technology itself has advanced rapidly. But he argues the investor question is the one that has not improved enough, and he points to the cost side. "In a lot of ways, companies are losing more money today implementing this technology than they were two years ago," he said, with the “hill” to climb getting steeper because spending has risen.

That’s the heart of his skepticism: enterprise ROI is still unanswered. Covello frames it as a simple test. "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." Goldman’s Global Institute, which takes a more optimistic long-term view through co-head George Lee, largely agrees the math is hard. Lee estimated that $7 trillion to $8 trillion will ultimately be spent on AI infrastructure, and that simply disrupting existing profit pools will not generate enough payback. In Lee’s view, net new economic activity is required, which puts even more pressure on measurable returns inside operating businesses.

The evidence on returns is not exactly comforting. Fortune cites widely discussed MIT research finding 95% of organizations reporting zero return on AI pilots. It also points to a 2025 EY survey where 99% of companies reported financial losses tied to AI-related risks, averaging $4.4 million per company. And at Fortune’s COO Summit earlier this week, executives described the gap between AI experiments and business payoff. Cognizant reported that 93% of jobs are already being disrupted by AI, six years ahead of its own 2023 projections, yet productivity gains that were expected have not shown up. Their researchers called this an "activation gap." The translation for operators is brutal: organizations may deploy pilots, but the operational “activation” required to convert models into outcomes is lagging.

Even inside teams that lean into AI, the human layer is proving stubborn. Fortune reports that Cisco’s Francine Katsoudas, EVP and Chief People Officer, said trust within teams using AI most intensely began to drop about 9 months in. That’s a signal that deployment is not just a software rollout. It is a change-management and incentives problem, one that can quietly throttle ROI even when the tools work.

Covello also highlights a supply-chain paradox that changes who benefits from spending. Hyperscalers like Amazon, Microsoft, Google, and Meta have not just maintained AI capex despite stock underperformance, they have increased it. Covello had predicted that sustained underperformance would trigger spending discipline, but it didn’t. Instead, he describes "FOMO at every level of the supply chain," where each player fears being left behind if someone else cracks the economic code first. The result, he argues, is an inversion where semiconductor companies, led by Nvidia, capture nearly all the economic value, while companies above them in the chain bleed cash. He says that dynamic is historically unprecedented, noting that in prior cycles semiconductor companies thrive when customers thrive, but now semis are profiting at the expense of others. Covello’s own view reportedly shifted as a result: he now favors hyperscaler stocks over semiconductors, arguing hyperscalers win in two of three likely scenarios, while only pure status quo keeps semis printing indefinitely.

Finally, Covello sees pressure building inside enterprises, not just in markets. He notes that workers are not receiving the benefits C-suite leadership expects. Part of the bottleneck is data readiness. "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," he said. That leads to a second economic challenge: agents sitting on top of data that cannot yet support agentic workflows. Goldman’s George Lee adds another drag coefficient for incumbents: legacy systems, entrenched workflows, and resistance to change slow down productivity gains at scale. The productivity leap is real, but it is primarily visible in companies built from scratch for this technology. For everyone else, it takes longer, and the “longer” window is exactly what IPO hype might not get to enjoy.