June 24, 2026—Bitcoin fell below the $60,000 mark, briefly touching $59,023 and hitting its lowest point since October 2024. On the same day, Nvidia (NVDA) closed at $198.91, down 0.56%; AMD closed at $519.85, down 5.76%. The previous trading day (June 23), the Nasdaq had already plunged 2.21%, with Google down over 5%, Amazon down 4.75%, Microsoft down 3.18%, and Meta down 2.32%.
What’s driving the market panic?
The answer isn’t complicated. Financial reports reveal that Google, Amazon, Microsoft, and Meta—the four hyperscale cloud giants—are projected to ramp up their combined capital expenditures to $725 billion in 2026, a 77% increase from $410 billion in 2025. Meanwhile, the hourly rental price for Nvidia’s B200 chip dropped from $6.11 on May 30 to $4.22 on June 22—a 31% decline in less than a month.
Capital spending is surging, while compute rental prices are tumbling. Is the return on investment (ROI) for AI chips really adding up? This article aims to break down the logic behind this debate, using the latest data.
Where Does the $725 Billion Go? Who’s Spending, and on What?
To understand the ROI debate, we first need to examine the structure of capital expenditures.
Goldman Sachs’ updated forecast from June 2026 shows that the four hyperscale data center operators—Alphabet (Google), Amazon, Microsoft, and Meta—will collectively spend $725 billion in capital expenditures in 2026. Breaking it down: Amazon is expected to spend around $200 billion, Microsoft about $190 billion, Google between $175 and $185 billion, and Meta between $115 and $135 billion.
What does this number mean? In just the past six months, market expectations for 2026 cloud provider capex have jumped nearly 80%. $725 billion exceeds the total size of the global semiconductor market in 2025—according to the World Semiconductor Trade Statistics (WSTS), the global semiconductor market is projected to reach $1.5112 trillion in 2026—meaning these four companies’ AI capex will account for nearly half the entire global semiconductor market.
This spending generally flows into three tiers: upstream chip procurement (Nvidia GPUs, AMD accelerators, custom ASICs, etc.); midstream data center infrastructure (land, buildings, power, cooling systems); and downstream network equipment and software ecosystems (InfiniBand, Ethernet, CUDA, etc.). Bernstein Research notes that a rise in HBM (high-bandwidth memory) prices alone could increase hyperscale cloud AI capex by about 30%.
But the core issue isn’t "where the money goes"—it’s whether the money can be earned back.
Single-Digit ROIC: Short Sellers’ Warnings and the Math Behind It
Wall Street legend and famed short seller Jim Chanos provided a concrete figure at a June 2026 seminar: the expected pre-tax return on invested capital (ROIC) for current compute infrastructure is only 5% to 8%.
Chanos’ logic is straightforward. He points out a significant "financial mismatch" in the current AI industry chain: companies selling chips and data center equipment are immediately recognizing revenue and profits, while the cloud providers buying this equipment are capitalizing the costs. Once these assets go live and begin depreciating, the hit to profits will be substantial.
He draws a parallel to the internet bubble of 1998–2000. Back then, S&P 500 operating profits grew 30% in two years, but when order books collapsed in 2001 and depreciation costs kept mounting, S&P 500 profits plunged 40%.
Chanos further highlights the essence of the compute rental model: if you buy chips from Nvidia, rent someone else’s data center, and then sublease compute power to Microsoft or Google, you’re essentially an equipment rental company—not a tech company.
This assessment aligns with market data. Nvidia B200 rental prices have dropped 31% in a month, and overall AI server rental costs continue to decline. If scarcity fades, the rationale for sustained capex weakens.
Jensen Huang’s Response: "Useful AI" Has Arrived
Facing market skepticism, Nvidia founder and CEO Jensen Huang delivered a direct response at the company’s annual shareholder meeting on June 24.
He led with financials: Nvidia’s revenue for fiscal year 2026 surged 65% to $216 billion, with operating cash flow reaching $103 billion. Data center revenue alone grew 68% to $194 billion. Tens of thousands of Blackwell-architecture GPUs have already been deployed by hyperscale cloud providers and model developers.
Huang’s core argument: "Useful AI" has arrived—and it’s already profitable. He believes the market’s ROI question for AI investment "already has an answer." In his view, AI is driving the largest industry reset in computing in 60 years: from humans writing software and computers executing instructions, to computers understanding, reasoning, planning, and performing real work. AI data centers are no longer just storage hubs—they’re "token factories."
This logic rests on a key premise: if AI can generate real economic value (boosting productivity, replacing labor, creating new business models), then infrastructure investment has a foundation for returns. The problem is, this "if" still lacks large-scale, verifiable financial data.
Compute Rental Prices: The Most Honest Market Signal
Amid all the debate, compute rental prices may be the most objective reference point.
According to GPU compute price tracking platform Ornn, the hourly rental price for Nvidia’s B200 peaked at $6.11 on May 30—the highest in nearly three months—then steadily declined to $4.22 by June 22, a 31% drop.
This trend sends several signals. In the short term, supply is rapidly catching up with demand—as Blackwell GPUs ship in volume, compute shortages are easing. In the medium term, if rental prices keep falling, "AI cloud" providers whose business models rely on leasing compute power will face margin compression. In the long run, lower prices could spur more demand at the application layer, creating a "price cut–volume growth" positive cycle—but this remains to be seen.
Notably, compute rental price trends have moved in tandem with tech stock performance. On June 23, the Nasdaq fell 2.21% and Nvidia dropped 4.13%. The market seems to be answering the same question with price signals: Is compute power really still that scarce?
Custom Chips: Hyperscalers’ Path to "Decoupling"
Facing tight GPU supply and soaring procurement costs, hyperscale cloud providers are accelerating their custom chip strategies.
A recent J.P. Morgan report predicts that shipments of AI custom chips (ASICs/XPUs) will surpass GPUs for the first time in 2027. In 2026, the AI ASIC market is expected to reach $60–70 billion, with a compound annual growth rate of 40% to 50%. ASICs are projected to account for 42% of all AI chip shipments in 2026, rising to 53% in 2027.
Broadcom currently holds 80% to 85% of the high-end ASIC market. Google’s TPU, Amazon’s Trainium/Inferentia, and Meta’s MTIA series have all entered large-scale deployment.
The core logic behind custom chips: control the supply chain, lower the unit cost of compute, and reduce reliance on a single supplier. Meta expects 2026 capex of $115–135 billion—nearly double the previous year—with its custom MTIA chip promising a 44% cost reduction.
However, custom chips also carry significant sunk cost risks. Chip design, tape-out, validation, and software ecosystem adaptation all require massive upfront investment, and the technology is evolving rapidly—Meta’s MTIA series plans a new generation every six months. If the AI investment cycle ends prematurely, these investments may not be recouped.
Conclusion: The Real Issue with ROI Is a Mismatch in Timing
Returning to the original question: Does the ROI on AI chips really add up?
Based on current data, the answer isn’t a simple "yes" or "no"—it’s a matter of timing mismatch.
Upstream suppliers (Nvidia, TSMC, Broadcom, etc.) are posting record revenues and profits. Nvidia’s FY2026 revenue hit $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 under pressure from capital expenditures. The $725 billion in annual spending will need to be recouped over the coming years through AI service revenues, improved ad efficiency, enterprise software subscriptions, and more. Chanos’ 5% to 8% ROIC estimate, together with Bernstein’s assertion that "cost rebalancing is inevitable," form a mutually reinforcing outlook.
The market is voting with its wallet. On June 24, Bitcoin fell below $60,000 as capital rotated from crypto into AI-related tech stocks; just a day later, tech stocks themselves faced heavy selling. This volatility in asset prices is the market’s truest expression of ROI uncertainty.
The debate over AI chip ROI won’t be settled anytime soon. It will be revisited and recalibrated with every update to capex figures, compute rental prices, cloud provider earnings, and chip shipment volumes. For investors, one thing is certain: the industry is shifting from "faith-driven" to "data-driven."
FAQ
Q1: What is the total capital expenditure for the four hyperscale cloud providers in 2026?
Google, Amazon, Microsoft, and Meta are projected to spend a combined $725 billion in capital expenditures in 2026, a 77% increase from $410 billion in 2025. Amazon is expected to spend about $200 billion, Microsoft $190 billion, Google $175–185 billion, and Meta $115–135 billion.
Q2: What is the AI infrastructure ROI calculated by Chanos?
Based on current transaction details, Jim Chanos estimates the expected pre-tax ROIC for compute infrastructure at just 5% to 8%, all in the single digits. He notes that if this is the best they can do even during a chip shortage, he remains highly skeptical about downstream profitability.
Q3: What recent changes have occurred in the rental price of Nvidia’s B200 chip?
According to GPU compute price tracker Ornn, the hourly rental price for the B200 fell from $6.11 on May 30 to $4.22 on June 22—a 31% drop in less than a month. The market interprets this trend as a sign that compute supply is rapidly catching up with demand.
Q4: When will ASIC chips surpass GPUs in market share?
J.P. Morgan predicts that shipments of AI custom chips (ASICs/XPUs) will surpass GPUs for the first time in 2027. ASICs are expected to account for 42% of all AI chip shipments in 2026, rising to 53% in 2027. The AI ASIC market is projected to reach $60–70 billion in 2026.
Q5: How did Nvidia perform in fiscal year 2026?
Nvidia’s revenue reached $216 billion in FY2026, up 65% year-over-year, with operating cash flow of $103 billion. Data center revenue grew 68% to $194 billion. Tens of thousands of Blackwell-architecture GPUs have already been deployed by hyperscale cloud providers and model developers.
