In September 2026, the Nasdaq Composite closed last week at an all-time high. Yet the disagreement between bulls and bears over AI stocks is intensifying. Michael Burry—an investor famous for accurately predicting the 2008 subprime mortgage crisis—has shifted his short positions in AI-related stocks into put options. In his Substack newsletter, he says that weekend research has convinced him the AI bubble could burst earlier than he originally expected.
At the same time, however, Nvidia’s quarterly revenue is nearing $100 billion. Micron’s HBM revenue has topped $1 billion for two straight quarters. Palantir’s U.S. commercial business year-over-year growth rate has reached 149%. Goldman Sachs’ latest research framework shows that six hyperscalers’ AI compute capital expenditure from 2026 to 2027 will total about $1.73 trillion. To achieve a 15% return on invested capital, they need to generate roughly $1.42 trillion in cumulative revenue from 2028 to 2030. What do these figures actually mean? What else can still support valuations of AI stocks? This article breaks down the question across four dimensions: compute, storage, AI software commercialization, and capital expenditure investment returns.
Nvidia: Demand for compute remains the first pillar supporting valuation
In Nvidia’s fiscal second quarter of 2026 (ended July 2026), revenue hit $96.2 billion, up 106% year over year. Data center revenue was $89.0 billion, up 117% year over year. GAAP gross margin held at 75.0%, and earnings per share were $2.46, up 128%. The third-quarter revenue outlook is $108.0 billion; if met, it would set another record for a single quarter.
On valuation, according to Gate Stock quotes, as of the close on September 28 Eastern Time, Nvidia’s stock price was $228.86, with a market cap of about $5.52 trillion and a TTM price-to-earnings (P/E) ratio of about 28.93x. Notably, although the stock is up about 21% year to date, its forward P/E has fallen to 16.6x. The reason is that the market’s blended consensus forecast for earnings per share over the next 12 months rose by 83%, causing the forward P/E to drop sharply from 25.3x at the end of 2025.
Source: Gate Stock quotes
This set of data supports a key conclusion: Nvidia’s valuation expansion isn’t driven purely by sentiment. The decline in the P/E multiple comes from earnings growth outpacing the stock’s price gains, not from a pullback in the share price. In other words, valuation is being absorbed by profit growth. That is fundamentally different from how tech stocks were priced during the 2000 internet bubble using the "price-to-dreams" multiple.
The risk side also can’t be ignored. Goldman Sachs assumes a 5-year economic life for GPUs in Nvidia’s products for depreciation. If the actual life is only 3 years, then annual depreciation and amortization would rise significantly, directly eroding investment returns. In addition, Nvidia’s roughly $3 trillion in off-balance-sheet commitments (mainly compute procurement agreements and long-term supply contracts) are not captured in most valuation models.
Micron Technology: Earnings leverage in storage, plus cyclical risk
Micron’s fiscal third-quarter revenue was $41.46 billion, up about 346% year over year. Adjusted EPS was $25.11, up more than 12x. Adjusted gross margin reached 84.9%. Among this, cloud storage revenue was $13.77 billion, up more than 300% year over year. HBM revenue topped $1 billion for the second consecutive quarter. Micron’s HBM4 ramp to mass production is already moving at twice the pace of the prior-generation HBM3E. Management said clearly that Micron’s 2026 HBM capacity is essentially sold out, and that the tight supply-demand landscape is expected to persist into—and beyond—2027.
According to Gate Stock quotes, as of September 28, Micron’s stock price was $1,053.98, with a market cap of about $1.19 trillion and a TTM P/E ratio of about 23.86x. The forward P/E based on future earnings expectations is about 7.32x. This valuation level isn’t high for storage chips, a strongly cyclical industry. The core reason is that the market’s concern about the cycle topping remains persistent.
Source: Gate Stock quotes
In a September update on an HBM industry research note, JPMorgan said that even if HBM specifications are downgraded, the compound annual growth rate in HBM bit demand from 2026 to 2028 would still be 63%. The industry remains in a supply-constrained state, so the upward trend in pricing is expected to continue. JPMorgan expects the average selling price of HBM to rise 54% year over year in 2027, and to continue rising 25% in 2028. The global storage market is expected to grow from about $969 billion in 2026 to $1.44 trillion in 2027, and to $1.82 trillion in 2028.
But the cyclical nature of the storage industry won’t disappear simply because AI demand exists. If the growth rate of AI training demand slows, or if GPU shipments come in below expectations, the HBM supply-demand balance could reverse within two quarters. JPMorgan also noted that the year-over-year growth in shipment volumes of ASICs the industry designs in-house can reach 102%. By 2027, ASIC demand’s share could rise to 48%, surpassing Nvidia to become the largest end consumer of HBM. This structural shift could have a profound impact on Micron’s customer concentration and pricing strategy.
Palantir: Valuation alignment for AI software commercialization
Palantir’s second-quarter 2026 revenue was $1.935 billion, up 93% year over year. Of that, U.S. commercial revenue grew 149% to $764 million. U.S. government revenue rose 90% to $809 million. U.S. business now accounts for about 81% of total revenue. The company raised its full-year revenue guidance to about $8.15 billion, implying growth of about 82%.
However, on valuation, Palantir’s controversy is far greater than that of Nvidia and Micron. According to Gate Stock quotes, as of September 28, Palantir’s stock price was $187.48, with a market cap of about $450.5 billion. Its TTM P/E is as high as 159.76x, and its forward P/E remains in the 85 to 100x range.
Source: Gate Stock quotes
Measured by PEG (the price-to-earnings ratio relative to the earnings growth rate), Palantir’s current PEG is about 1.87. That means the market has already priced in a substantial premium for its high growth over the coming years. Whether this valuation can hold depends on two key variables: first, whether the growth rate of U.S. commercial business can stay above 100%; second, whether the renewal rate of government contracts and the total contract value can keep climbing. In the second quarter, the company signed 220 contracts worth $1 million or more. Total contract value rose 49% year over year to $3.373 billion, providing some support for near-term revenue visibility.
That said, AI software is also facing competitive pressure from foundation model companies. Palantir positions itself as an "AI sovereignty" platform and highlights differentiated capabilities—data never leaves the premises and models are auditable. That creates a moat in government and defense scenarios, but the competitive landscape is still unclear in pure commercial settings.
Capital expenditure scale and investment returns: Macroscale constraints on valuation support
Ultimately, the micro-level earnings growth of AI stocks depends on whether capital expenditure at the macro level can stay sustainable. Goldman Sachs divides this AI compute buildout cycle into three stages: Stage 1 (2023 to 2025) totals about $633 billion in capital expenditure, or about $211 billion per year on average. Stage 2 (2026 to 2027) totals about $1.73 trillion, or about $863 billion per year on average. Stage 3 (2028 to 2030) is expected to reach as much as $4.14 trillion in total capex, or about $1.38 trillion per year.
This means that the capital expenditure supporting revenues for Nvidia and Micron is approaching the financial tolerance limits of cloud providers. Goldman Sachs’ logic is as follows: with a 15% annualized ROIC target, six companies need to generate about $129.5 billion in after-tax net operating profit each year. Add about $276 billion per year in depreciation and amortization from Stage 2 assets. Then Goldman estimates that cumulative revenue over the three years from 2028 to 2030 must reach about $142 billion (i.e., $1.42 trillion) to support the investment. That corresponds to roughly a 35% EBIT profit margin. Currently, the combined backlog of the three major public cloud providers totals about $1.69 trillion, up 152% year over year, which offers some backing for achieving the revenue target. Still, depreciation and amortization will become the core pressure on the profit side.
Also according to Goldman Sachs, the median forward P/E multiple for AI infrastructure stocks has compressed sharply—from 32x in April 2026 to 22x today. The market is pricing in the uncertainty around capital returns through valuation compression.
Conclusion
Valuation support for AI stocks isn’t a single-dimension issue. On the micro level, Nvidia’s earnings growth is absorbing valuation at a faster pace than the stock price is rising. Micron’s tight HBM supply-demand supports clear earnings leverage in storage. And Palantir’s U.S. commercial growth rate provides some fundamental backing for its high valuation. But on the macro level, Goldman Sachs’ estimated $1.73 trillion in capex implies a $1.42 trillion revenue return threshold, and the potential substitution effect from ASICs for GPUs—along with disputes about GPU depreciation life—create hard constraints on valuation support. Investors should focus less on whether "the bubble will burst," and more on whether "earnings growth can complete the digestion of valuation before the capex wave withdraws." This validation process will unfold gradually over the next 12 to 18 months.
FAQ
Q1: What is the fundamental difference between the valuation of AI stocks and the internet bubble era?
During the internet bubble era, many companies had no revenue and no profits, so valuations depended on user growth expectations. Today, Nvidia’s TTM P/E is about 29x, while its forward P/E is only about 16.6x. Micron’s forward P/E is about 7.3x. In both cases, there is real earnings support. The key risk centers on the sustainability of capital expenditure, not on whether corporate earnings are real.
Q2: How long can the HBM supply-demand tightness last?
JPMorgan expects the compound annual growth rate in HBM bit demand from 2026 to 2028 to be 63%, and the shortage state to continue through 2028. Micron management said capacity is already sold out, and the tight setup is expected to last beyond 2027. Still, investors should watch for the substitution effect of ASIC chips for GPUs, and whether ASIC’s share of HBM demand could rise to 48% in 2027.
Q3: Is Palantir’s P/E ratio too high?
Palantir’s TTM P/E is about 160x, and its forward P/E is about 85 to 100x—far higher than typical SaaS peers. The reasonableness of its valuation depends on whether U.S. commercial business can sustain growth above 100%. Current total contract value is up 49%, which provides some near-term visibility, but the forward competitive landscape remains unclear.
Q4: What does a slowdown in cloud providers’ capital expenditure mean for AI stocks?
Goldman Sachs estimates that AI capital expenditure will be about $1.73 trillion in 2026 to 2027. To achieve a 15% ROIC, it needs about $1.42 trillion in revenue from 2028 to 2030. If revenue falls short of expectations, the order growth rates for GPUs and HBM will likely slow in tandem. The market’s pricing logic could shift from "revenue growth" to "margins and free cash flow," putting valuation multiples under compression pressure.
Q5: What does Michael Burry’s put option strategy mean?
Burry converted his short positions in AI stocks into put options to get higher leverage at a lower cost. His Micron put options expire in June 2027, while the put options for Nvidia and Palantir are valid through September 2027. This strategy reflects a short-term bet on the AI capital expenditure return cycle, not a long-term rejection of the companies’ fundamentals.
