The AI hardware cycle is pivoting from token maximization to return on invested capital, and the shift is already showing up in cash flows, memory prices, and Washington policy debates.
返回
AI hardware cycle shifts from cash burning to ROI-focused scaling
The AI hardware cycle is pivoting from token maximization to return on invested capital, and the shift is already showing up in cash flows, memory prices, and Washington policy debates.
After two years of relentless capital spending, the AI hardware cycle is entering a new phase where cloud providers are scrutinizing returns, memory prices are decelerating, and the burden of funding expansion is shifting from cash flow to balance sheets.
"The question has moved from 'can they spend it' to 'what's the return and who pays for it,'" a semiconductor analyst who participated in the roadshow said, summarizing the tone of investor meetings.
North American cloud providers are expected to spend $7.7 trillion on capital expenditures in 2026, up 88% from 2025, and $1.02 trillion in 2027, according to consensus estimates. But Google's second-quarter 2026 earnings on July 24 crystallized investor anxiety: capital spending reached 115% of operating cash flow, pushing free cash flow negative for the first time since the company went public. The stock fell despite cloud revenue surging 82%.
The shift matters because the AI hardware trade has been built on ever-rising CapEx guidance. If investors start discounting spending that doesn't generate measurable returns, the valuation premium on capital-intensive names could compress. The next flashpoint is memory, where DRAM price momentum is fading and the industry's cash-rich manufacturers face pressure to fund the next wave of expansion.
Memory Prices Signal a Slowdown, Not a Collapse
The consensus on DRAM pricing has converged: the slope of increases is flattening. Traditional DRAM contract prices rose 90% to 95% quarter over quarter in the first quarter of 2026, slowed to 58% to 63% in the second quarter, and are expected to rise just 13% to 18% in the third quarter, according to TrendForce data. Prices are still at record levels, but the second derivative has turned negative.
The real debate is about 2027. Optimists expect HBM (high-bandwidth memory) to catch up in annual contract renegotiations, eliminating the current price inversion with DDR5. Pessimists worry that memory manufacturers will accelerate equipment purchases — including ASML's High-NA EUV lithography tools — bringing new supply online faster than demand can absorb it.
The divergence in investor questions captured the uncertainty: half asked whether memory has peaked, while the other half asked whether stocks trading at less than five times next year's earnings have bottomed.
The Financing Chain Shifts From Cash to Credit
The most underappreciated risk in the AI hardware cycle is who ultimately pays for it. Nvidia and Advanced Micro Devices have begun offering equity-linked financing to OpenAI and Anthropic in exchange for chip orders, effectively converting customer concentration risk into balance sheet exposure. Taiwan Semiconductor Manufacturing Co. has publicly ruled out participating in such arrangements.
The next test is memory manufacturers, which hold the deepest cash reserves in the semiconductor industry. If they begin extending financing to cloud customers, the funding burden for the expansion cycle will have migrated from cloud providers' operating cash flow to the most cyclically exposed balance sheets in the supply chain.
On the policy front, the U.S. Congress is advancing two bills that could reshape the competitive landscape. The RASA (Remote Access to Sensitive Assets) Act, which would require licenses for cross-border computing access, has passed the House and is pending in the Senate. The MATCH Act, which would align export controls with the Netherlands and Japan and extend them to immersion DUV lithography tools, remains in proposal stage. A U.S.-China AI dialogue is being prepared for September, according to Bloomberg.
For investors, the cycle is not ending — it is rotating. Semiconductor equipment stocks, led by ASML, offer the most consensus upside, with global wafer fabrication equipment spending projected to reach $2.4 trillion by 2028, nearly double 2025 levels. Terminal device makers such as Apple, which have been overlooked in the AI narrative, have outperformed some core AI names this year, raising the question of whether "anti-AI" assets offer relative value when the CapEx story loses momentum. The winners of the next phase will be those whose spending generates measurable returns — not just those who spend the most.
This article is for informational purposes only and does not constitute investment advice.
