Michael Cembalest's latest Eye on the Market1 lands at a remarkable moment: a leadership change at the world's most consequential technology company, a geopolitical reckoning years in the making, and an AI capital cycle still running at full speed. The Icarus Files is not a eulogy. It is a stress test.
The Fastest Rise in S&P 500 History
Since 1960, ten companies have held the distinction of being the largest in the S&P 500. Cisco held the title for three days. GE held it for years. Nvidia's climb was something different entirely. As Cembalest documents, the company "rose from 1% of the S&P 500 in December 2022 to 8.5% by May 2026," a velocity of ascent unmatched in the index's modern history.
Then came the fall from grace. Nvidia's CEO Jensen Huang, Cembalest writes, "flew too close to the sun and the company's leadership will now change." The repricing that followed was swift and severe: Nvidia "now trades at the lowest valuation among large AI stocks by a wide margin," including a price-to-earnings-growth (PEG) ratio of just 0.2x, far below peers and sectors.
The historical precedent is sobering. Most companies that climbed to the S&P 500's top spot eventually descended to a small fraction of their peak weight. Cisco hit a P/E of 200x and price-to-sales of roughly 35x at its zenith. GE's collapse was driven by "the black box shenanigans of GE Capital," which at its peak represented 60% of GE profits. IBM "fell victim to the personal computer and open-systems revolutions which disrupted its high-margin mainframe monopoly." The pattern is consistent: exceptional reach, then gravity.
The Microsoft and Apple Exception: Can Nvidia Follow?
Not every summit ends in a crash. Microsoft and Apple managed to sustain elevated S&P 500 market cap shares over long periods, bucking the historical norm. Cembalest sees a near-term parallel for Nvidia: the company "earned 83% of AI chip revenues from 2023-2025 with gross margins of 75% and operating margins of 66%." Free cash flow margins have been "sustained at extraordinarily high levels," while hyperscaler free cash flow has been eroding under a surge in debt-financed capital spending.
The semiconductor industry's long test, design, and build cycles mean that Nvidia's cash flow dynamics "should be maintained for some time." The financials remain extraordinary. The PEG ratio of 0.2x suggests the market may be heavily discounting future earnings growth. Cembalest notes that "Nvidia looks interesting for long term holders at 32% below its all-time high," but flags that the competitive landscape may be just as material a risk as the management transition.
The Competitive Threat No One Is Fully Pricing
The most consequential section of The Icarus Files concerns the accelerator landscape. Nvidia's historical 80-90% AI chip revenue share in 2024-2025 tells only part of the story. Among announced frontier lab compute agreements, Nvidia holds just a 28% share. The remaining 72% is distributed across AMD GPUs, Amazon Trainium, Google TPUs, Broadcom custom ASICs, Microsoft Maia, and Cerebras.
Cembalest points to two critical data points. First, AMD's MI455X "could rate higher" on inference performance than Nvidia's current offerings, partly due to "more high bandwidth memory content and higher peak memory bandwidth." A year ago, onboarding and running models on AMD GPUs took six to twelve months of expensive engineering. Using their latest AI models to do that engineering, frontier labs "can onboard new models on AMD GPUs in just 3 days." Nvidia's core argument, that its CUDA software ecosystem made switching uneconomic, has been partially eroded.
Second, when comparing Nvidia's GB300 against Google's TPU on the Chinese open weight model Qwen 3.5 397B, "across all speeds except the slowest ones, Google's TPU looks better." The economics, as frontier labs vote with their feet, appear to be shifting.
The China Problem: Geopolitics Meets Supply Chain
Nvidia's reliance on China is not incidental. It is structural. Chinese open weight models consume the majority of tokens tracked through OpenRouter and their capabilities continue to track improvements in closed frontier models. Beyond token volume, Cembalest maps China's end-of-2027 AI compute sources: approximately 36% already inside China, 32% accessed remotely from overseas data centers (predominantly in Southeast Asia), and the remainder through legal imports, smuggling, and domestic production.
The Remote Access Security Act, which passed the US House 369-22, would restrict Chinese companies from remotely using restricted US chips in overseas data centers. It has not come to a Senate vote. Meanwhile, Cembalest documents Huang's consistent posture: calling concerns about the Chinese Communist Party "lunacy," labeling "China hawk" a "badge of shame," and lobbying against export controls as early as 2022. The share of Nvidia revenue billed to customers with Singapore addresses "rose from 8% to 22%" following Biden-era export controls, and in July 2025, Nvidia "stopped breaking out the figures entirely."
As analyst Andrew Sharp writes, in a passage Cembalest quotes at length: "While reaping the unbelievable rewards of US markets, laws and customers, Nvidia has not really behaved like an American company that's answerable to the American government."
The Neocloud Fallout and AI Debt Markets
Nvidia's vendor financing model is a hidden vulnerability that few are discussing. Over the last three quarters, Nvidia "spent 58% of its operating cash flow on equity investments and M&A." Off-balance sheet long-term debt and lease liabilities "rose from $214 bn to $422 bn" from Q1 to Q2 2026, rising to $530 billion when land, power, and shell guarantees are included. The Bank for International Settlements flagged these circular relationships: AI-to-AI deals, circular and non-circular, represented "more than 50% of all investments in AI firms from 2021-2025."
If Nvidia's new leadership pulls back from vendor financing, neocloud companies including CoreWeave, Nebius, Applied Digital, and IREN face meaningful additional downside. Yields on AI infrastructure debt, both high yield and investment grade, "have already been rising sharply this year." The cost of capital for AI companies is already climbing without further reduction in Nvidia's ecosystem support.
The Big Picture Still Holds
Despite the turbulence, Cembalest urges perspective. The AI capital cycle "is still in the middle of the largest capital spending boom since the US railroads in the 1880's." Legacy Nvidia GPUs continue to command rising rental prices. Taiwanese and South Korean chip exports are climbing. US data center vacancy rates remain below 1%. North American voter revolts against data center proliferation may paradoxically benefit Nvidia: scarcer capacity would pressure owners to "continually upgrade to better and more efficient chips in order to increase token output and revenues."
Five Key Takeaways for Advisors and Investors
- The leadership transition matters, but the valuation does too. At a PEG of 0.2x and 32% below its all-time high, Nvidia's risk/reward calculus is different from what it was at its peak. Long-term investors are not wrong to look closely, but they must price in both management risk and competitive risk simultaneously.
- Nvidia's moat is narrowing at the hardware level. AMD's MI455X and Google's TPUs are demonstrably competitive on inference performance. CUDA's switching-cost advantage is eroding faster than consensus expected. Advisors should treat Nvidia's market share projections with more skepticism than they did a year ago.
- China exposure is a regulatory overhang, not a sideshow. The RASA bill, Congressional inquiries into Megaspeed, and ongoing export control enforcement are live risks. If Chinese access to US chips is materially curtailed, Nvidia loses a core demand base and Chinese firms have greater incentive to accelerate domestic alternatives like Huawei's Ascend NPUs.
- Neocloud and AI infrastructure debt are second-order risks. The circular financing architecture Nvidia built is fragile. If vendor financing contracts under new leadership, neocloud valuations, already under pressure, face further downside, and AI debt markets are not fully priced for that scenario.
- The underlying AI demand cycle remains intact. Rising GPU rental prices, record chip exports, and near-zero data center vacancies confirm that AI infrastructure spending is still accelerating. The Nvidia story has become more complicated; the AI story has not.
Footnote:
1 Cembalest, Michael. "The Icarus Files: What's Next for Nvidia and AI." Eye on the Market, J.P. Morgan Asset Management, 14 Nov. 2026, https://cdn.jpmorganfunds.com/content/dam/jpm-am-aem/global/en/insights/eye-on-the-market/the-icarus-files-amv.pdf.