Morgan Stanley and Goldman Sachs flag a growing gap between AI hardware forecasts and what the semiconductor industry can realistically deliver
The AI chip market is having a moment that feels uncomfortably familiar. Demand is surging, valuations are climbing, and Wall Street is revising forecasts upward faster than chip fabs can spin up production lines. The problem, according to analysts at Morgan Stanley and Goldman Sachs, is that the math stops working somewhere between the PowerPoint deck and the factory floor.
Both firms have raised concerns that AI chip revenue forecasts for 2026 and 2027 have outpaced what realistic capacity and demand trajectories can actually support.
How we got here #
The current frenzy traces directly back to the large language model breakthroughs of 2022 and 2023. ChatGPT arrived, cloud providers started placing GPU orders at scale, and NVIDIA’s H100 and A100 GPU series, built on the CUDA software ecosystem, became the de facto standard for training and running large AI models. NVIDIA’s data center segment became the primary engine of growth across the entire AI hardware category.
Corporate guidance followed the hype upward. Consecutive quarters brought upward revisions, and market valuations for leading AI hardware companies got rebuilt around an assumption of sustained annual growth exceeding 30%. For semiconductors, a sector with a well-documented history of boom-and-bust cycles, it is the kind of number that makes veteran analysts reach for antacids.
Here is the thing about 30% annual growth: it compounds fast. A company generating $10 billion in revenue today would need to be doing roughly $27 billion in four years to hit that mark. Sustaining that trajectory requires not just demand, but also manufacturing capacity, a functional supply chain, and customers beyond the small club of hyperscalers who can actually write those checks.
The supply side has its own problems #
Even if demand holds up, production lead times for advanced AI chips are currently stretching to between 6 and 12 months at contract manufacturers, with TSMC being the critical bottleneck in that chain. TSMC manufactures chips for NVIDIA, Apple, AMD, and most of the other names that matter. When lead times push past half a year, it creates conditions where supply and demand can fall badly out of sync.
The crypto mining boom offers a cautionary parallel. Between 2020 and 2022, GPU demand from crypto miners sent orders soaring. Manufacturers ramped production. Then mining profitability collapsed, demand evaporated, and the industry spent the better part of a year working through an inventory hangover. Prices dropped, margins compressed, and companies that had bet on sustained demand found themselves holding excess stock.
Margins are already under pressure. Recent quarterly results across major chipmakers show growth in AI-related orders, but research and development spending is climbing, fabrication investments are increasing, and the race to develop next-generation architectures means the capital expenditure cycle has no obvious end point.
What this means for investors watching the sector #
A significant portion of current AI chip demand flows from a handful of hyperscale customers. If even one or two of those buyers s, delays, or redirects spending, the demand picture changes materially.
Broader enterprise adoption is still developing. Enterprises are experimenting with AI. Fewer are deploying it at the scale that generates meaningful chip demand. The transition from experimentation to production deployment takes time, and the timeline is uncertain enough that baking aggressive assumptions into 2026 and 2027 forecasts looks like a stretch.
Valuations built on 30%-plus annual growth leave little room for disappointment. The semiconductor sector has historically been valued on cyclically adjusted multiples precisely because investors learned, repeatedly, that peak earnings are not permanent earnings.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our