The Bear Market Whispers: Why Goldman Sachs Is Preparing for a Brutal Q3 Correction

On a busy morning, you can see the typical controlled intensity of the trading floors at 200 West Street: analysts crouched over two monitors, screens scrolling with numbers that most people will never comprehend, and the entire global finance apparatus operating at its usual low roar. Nothing particularly concerning. However, a tension that hasn’t quite made headlines is building somewhere inside that building and in the research notes that are discreetly being distributed to institutional clients. The summer is making Goldman Sachs anxious. Additionally, it’s usually worthwhile to pay attention when Goldman becomes anxious.

Although the mechanics are complex, the main issue is not. The value of US stocks is almost at its highest point in the last 20 years. For over nine months, there hasn’t been a significant decline in the markets. This is the kind of prolonged, uninterrupted rise that typically ends with something much more abrupt rather than a smooth decline. As early as October 2025, Goldman CEO David Solomon voiced concerns, and the company’s overall 2026 forecast has been consistent ever since: a complete bear market isn’t the most likely scenario, but a significant correction isn’t only conceivable—it might be past due. A dollar at multi-year lows, shifting trade alliances, and uncertainty surrounding the Federal Reserve could all contribute to the match.

Whether people want to acknowledge it or not, there’s a feeling that the AI investment story is at the heart of it all. The figures are truly astounding. Together, the five largest tech companies are expected to invest about $371 billion in data centers and AI infrastructure this year. According to McKinsey, just meeting the demand for AI will cost the world $5.2 trillion by 2030. To put things in perspective, that is more than seven times the total cost of constructing the US Interstate Highway System. There is genuine ambition in those figures. Whether the revenue can possibly catch up in time to support them is the question.

Bain estimated that Big Tech would require an extra $2 trillion in yearly income by 2030 to pay for data center expenses, and even in the best-case scenario, there would be a $800 billion shortfall. That disparity is huge, and it’s the kind of figure that usually goes unnoticed in analyst footnotes until it does. According to estimates, more than $30 billion in AI spending had already been shown to yield a pitiful return on investment by Q4 2025. According to US Census Bureau data, enterprise adoption of AI at large firms—those with 250 or more employees—actually decreased from 14% to 12% between mid-2025 and late 2025. The infrastructure is being constructed more quickly than the number of new clients.

The dotcom parallels are becoming more and more obvious. In 2000, Lucent and Nortel provided vendor financing to clients who used the funds to purchase equipment that they otherwise couldn’t afford. This loop appeared to be increasing revenue until it didn’t. Similar mechanisms are being observed by some analysts in the current AI ecosystem. Instead of selling GPUs, Nvidia will lease them as part of its $100 billion investment partnership with OpenAI. In exchange, Nvidia will receive equity in a $500 billion company while anticipating $5 billion in annual losses.

According to NewStreet Research, the structure generates $35 billion in GPU purchases or leases for every $10 billion Nvidia invests, which accounts for more than 25% of Nvidia’s yearly revenue. According to Jay Goldberg of Seaport Global Securities, the deal has “a whiff of circular financing.” The names Lucent and Nortel came to mind when Peter Boockvar, a veteran of the dotcom era, heard the details. For the record, those businesses collapsed spectacularly.

This does not imply that the technology is fraudulent. It isn’t. According to Goldman’s own long-term forecasts, generative AI could increase productivity by 1.5 percent annually over the next ten years and contribute about 7 percent to the world GDP.

Nvidia and Palantir are trading at about 40 and 69 times sales, respectively. These are undoubtedly bubble-territory multiples, but the underlying technology they are offering is truly potent in ways that the dotcom companies were never. Everything was altered by the internet. Along the way, it bankrupted hundreds of businesses, destroyed $5 trillion in market value, and left investors who purchased at peak valuations waiting ten years to break even. The market is currently in an uncomfortable position because both of these things were true at the same time.

The end of AI is not anticipated by analysts and models that converge on Q2–Q3 2026 as the likely correction window. They are forecasting what will happen when it is no longer possible to cover up the discrepancy between revenue generated and capital deployed with narrative. Disappointments with earnings mount. Business clients who conducted pilots covertly allowed them to expire without being renewed.

Influencers who built their careers on optimism about AI begin to hedge their language before vanishing into the future. Goldman is keeping an eye out for a correction that is more likely to be painful than catastrophic, a repricing rather than a collapse. However, painful is still painful, especially for those who bought into the idea that the story would continue to develop throughout the summer.

Even the cautious voices aren’t really contesting the long-term picture, which makes this moment especially peculiar. Goldman is still optimistic about the economic potential of AI. Solomon continues to use transformative language when discussing the technology. The question of whether the market has priced in a journey that will take significantly longer and cost significantly more than the current enthusiasm suggests is what is causing the hesitation, not whether the destination is real. This year’s most important financial story is probably the conflict between true technological power and truly stretched valuations. Additionally, Goldman Sachs appears to be keeping an eye on time, at least with regard to this specific question.