AI Chip ROI Controversy Explained: $725 Billion in Capital Expenditure—How Do Hyperscalers Validate Return on Investment?

On June 24, 2026, Bitcoin fell below the $60,000 mark, briefly touching $59,023 during the session, a new low since October 2024. On the same day, Nvidia (NVDA) closed at $198.91, down 0.56%; AMD closed at $519.85, down 5.76%. On the previous trading day (June 23), the Nasdaq had already dropped 2.21%, Google fell over 5%, Amazon fell 4.75%, Microsoft fell 3.18%, and Meta fell 2.32%.

What is the market afraid of?

The answer is not complicated. Earnings data shows that the combined 2026 capital expenditures of the four hyperscale cloud providers—Google, Amazon, Microsoft, and Meta—rose to $725 billion, a 77% increase year-over-year from $410 billion in 2025. Meanwhile, the hourly rental price of Nvidia's B200 chip fell from $6.11 on May 30 to $4.22 on June 22, a drop of 31% in less than a month.

Capital expenditures are surging, while compute rental prices are plummeting. Can the return on investment (ROI) for AI chips actually be justified? This article attempts to deconstruct the logic behind this controversy from a data perspective.

The $725 Billion Destination: Who Is Spending and Where

To understand the ROI debate, we first need to see the structure of capital expenditures.

Goldman Sachs' updated forecast released in June 2026 shows that the four major hyperscale data center operators—Alphabet (Google), Amazon, Microsoft, and Meta—will have total capital expenditures of $725 billion in 2026. Broken down by company: Amazon approximately $200 billion, Microsoft approximately $190 billion, Google approximately $175 billion to $185 billion, and Meta approximately $115 billion to $135 billion.

What does this number mean? In just the past six months or so, the market's expectations for cloud providers' 2026 capital expenditures have increased by nearly 80%. $725 billion exceeds the total size of the global semiconductor market in 2025—the World Semiconductor Trade Statistics (WSTS) predicts the global semiconductor market will reach $1.5112 trillion in 2026—meaning the AI capital expenditures of these four companies alone are close to half of the global semiconductor market.

The flow of these funds can be roughly divided into three layers: upstream is chip procurement (Nvidia GPUs, AMD accelerators, proprietary ASICs, etc.); midstream is data center infrastructure (land, buildings, power, cooling systems); downstream is networking equipment and software ecosystems (InfiniBand, Ethernet, CUDA ecosystem, etc.). Bernstein Research points out that the rise in HBM (High Bandwidth Memory) prices alone could increase hyperscale cloud providers' AI capital expenditures by approximately 30% overall.

But the core issue is not "where the money is spent," but "whether the money can be earned back."

Single-Digit ROIC: The Bear's Warning and Accounting Logic

Legendary Wall Street short seller Jim Chanos gave a specific number at a seminar in June 2026: the expected pre-tax return on invested capital (ROIC) for compute infrastructure is only 5% to 8%.

Chanos's logic is straightforward. He points out that there is a huge "financial mismatch" in the current AI supply chain: companies selling chips and data center equipment immediately recognize revenue and profits, while the cloud providers buying these assets capitalize the costs. Once these assets go online and begin depreciating, the impact on profits will be enormous.

He compares current AI infrastructure investment to the internet bubble period from 1998 to 2000. At that time, S&P 500 operating profits grew 30% in two years, but when the order book collapsed in 2001 and depreciation costs continued to show, S&P 500 profits plunged 40%.

Chanos further highlights the essence of the compute rental model: if you buy chips from Nvidia, rent data centers from others, and then sublease compute power to Microsoft or Google, you are essentially an equipment leasing company, not a tech company.

This judgment echoes market data to some extent. The rental price of Nvidia's B200 chip fell 31% in one month, while the overall rental cost of AI servers shows a continuous downward trend. If scarcity disappears, the logical foundation supporting sustained capital expenditure will be weakened.

Jensen Huang's Response: "Useful AI" Has Arrived

Facing market skepticism, Nvidia founder and CEO Jensen Huang gave a direct response at the annual shareholder meeting on June 24.

He first used financial data: Nvidia's fiscal 2026 revenue grew 65% to $216 billion, with operating cash flow reaching $103 billion. Among this, data center revenue grew 68% to $194 billion. Hundreds of thousands of Blackwell architecture GPUs have been deployed cumulatively among hyperscale cloud providers and model developers.

Huang's core argument is: "Useful AI" has arrived and is already generating money. He believes the market's question about AI investment returns "already has an answer." In his view, AI is driving the most significant industry reset in computing in 60 years—shifting from humans writing software and computers executing instructions, to computers understanding, reasoning, planning, and completing actual work. AI data centers are no longer traditional data storage centers but "factories producing tokens."

The logical premise of this argument is: if AI can create real economic value (improving productivity, replacing human labor, creating new business models), then infrastructure investment has a basis for returns. The problem is that this "if" currently lacks large-scale, verifiable financial data support.

Compute Rental Prices: The Most Honest Market Signal

Among all controversies, compute rental prices may be the most objective reference indicator.

According to data from GPU price monitoring platform Ornn, the hourly rental price of Nvidia's B200 rose to $6.11 on May 30, a three-month high; since then, it has fallen continuously, dropping to $4.22 as of June 22, a decline of 31%.

This trend sends several signals. In the short term, supply is rapidly catching up with demand—as Blackwell architecture GPUs ship in large volumes, compute shortages are easing. In the medium term, if rental prices continue to decline, "AI cloud" service providers that rely on compute leasing as their core business model will face margin compression. In the long term, falling prices may stimulate more application-layer demand, forming a positive cycle of "lower prices, higher volume"—but this requires time to verify.

Notably, compute rental price trends have formed a kind of resonance with tech stock performance. On June 23, the Nasdaq fell 2.21%, and Nvidia fell 4.13%. The market seems to be using price signals to answer the same question: Is compute power still that scarce?

In-House Chips: The Hyperscalers' Path to Decoupling

Facing the dual pressures of GPU supply tightness and high procurement costs, hyperscale cloud providers are accelerating their in-house chip strategies.

A recent report from JPMorgan predicts that shipments of AI custom chips (ASIC/XPU) will surpass GPUs for the first time in 2027. In 2026, the AI ASIC market is expected to reach $60 billion to $70 billion, with a compound annual growth rate of 40% to 50%. ASICs are expected to account for 42% of all AI chip shipments in 2026, rising to 53% in 2027.

Broadcom currently holds an 80% to 85% share in the high-end ASIC market. Google's TPU, Amazon's Trainium/Inferentia, and Meta's MTIA series have all entered large-scale deployment phases.

The core logic behind in-house chips is: controlling the supply chain, reducing per-unit compute costs, and reducing dependence on a single supplier. Meta expects capital expenditures of $115 billion to $135 billion in 2026, nearly double the previous year. Its in-house MTIA chip promises a 44% cost reduction.

However, in-house chips also face significant sunk cost risks. Chip design, tape-out, verification, and software ecosystem adaptation all require huge upfront investments, and technology iteration is extremely fast—Meta's MTIA series plans to update every six months. If the AI investment cycle ends prematurely, these investments may not be recovered.

Conclusion: The Essence of the ROI Debate Is a Timing Mismatch

Back to the original question: Can the ROI of AI chips be justified?

Based on existing data, the answer may not be a simple "yes" or "no," but a timing mismatch issue.

Upstream suppliers (Nvidia, TSMC, Broadcom, etc.) are recognizing record revenue and profits. Nvidia's fiscal 2026 revenue was $216 billion, up 65% year-over-year. TSMC's AI semiconductor revenue share is expected to rise from about 15% in 2024 to over 30% in 2026.

Downstream cloud providers are undergoing a capital expenditure stress test. The annual spending of $725 billion needs to be recovered over the next few years through AI service revenue, advertising efficiency improvements, enterprise software subscriptions, and more. Chanos's estimate of 5% to 8% ROIC, combined with Bernstein's assessment that "cost rebalancing is inevitable," forms a set of mutually reinforcing judgments.

The market is voting with prices. On June 24, Bitcoin fell below $60,000, with funds shifting from cryptocurrencies to AI-related tech stocks; just one day later, tech stocks themselves faced heavy selling. This fluctuation in asset prices is itself the market's most genuine expression of ROI uncertainty.

The ROI debate on AI chips will not end soon. It will be re-examined and recalibrated with every update on capital expenditure data, compute rental prices, cloud provider earnings, and chip shipments. For investors, the only certainty is that this industry is moving from "faith-driven" to "data-driven."

FAQ

Q1: What is the total 2026 capital expenditure of the four hyperscale cloud providers?

The combined 2026 capital expenditures of Google, Amazon, Microsoft, and Meta total approximately $725 billion, a 77% increase from $410 billion in 2025. Among them, Amazon is about $200 billion, Microsoft is about $190 billion, Google is about $175 billion to $185 billion, and Meta is about $115 billion to $135 billion.

Q2: What is the return on investment for AI infrastructure calculated by Chanos?

Based on current transaction details, Jim Chanos estimates that the expected pre-tax ROIC for compute infrastructure is only 5% to 8%, all single digits. He points out that if even during this chip shortage period it can only achieve this level, he is highly skeptical of the profitability of downstream segments.

Q3: What has happened to the rental price of Nvidia's B200 chip recently?

According to data from GPU price monitoring platform Ornn, the hourly rental price of the B200 fell from $6.11 on May 30 to $4.22 on June 22, a drop of 31% in less than a month. This trend is interpreted by the market as a signal that compute supply is rapidly catching up with demand.

Q4: When will ASIC chips surpass GPUs in market share?

JPMorgan predicts that shipments of AI custom chips (ASIC/XPU) will surpass GPUs for the first time in 2027. In 2026, ASICs are expected to account for 42% of all AI chip shipments, rising to 53% in 2027. The 2026 AI ASIC market size is expected to reach $60 billion to $70 billion.

Q5: How did Nvidia perform in fiscal 2026?

Nvidia's fiscal 2026 revenue reached $216 billion, up 65% year-over-year, with operating cash flow of $103 billion. Among this, data center revenue grew 68% to $194 billion. Hundreds of thousands of Blackwell architecture GPUs have been deployed cumulatively among hyperscale cloud providers and model developers.