MW Nvidia can deliver chips - but it can't buy Big Tech out of its credit and power-grid crisis
By Jurica Dujmovic
Quarterly earnings won't fix a chaotic trade war with China, climbing credit premiums or AI infrastructure limits
Nvidia, led by CEO Jensen Huang, reported no China data-center compute revenue in its fiscal first quarter and is not assuming any in its current-quarter outlook either, with uncertainty over whether Beijing will allow any imports at all.
Every component of the AI buildout - power, silicon, steel, water - is in short supply and getting more expensive.
The artificial-intelligence investment thesis has always rested on a single claim: The demand is real, the chips are scarce and the cash flows will follow.
On the demand side, the evidence holds. The four largest hyperscalers are tracking toward roughly $700 billion in combined capital expenditure for 2026 - up more than 60% from 2025's already record levels. Nvidia (NVDA) CEO Jensen Huang has put the company's Blackwell and Vera Rubin opportunity at a combined $1 trillion through 2027. Nvidia's fiscal first-quarter results, announced after the U.S. market close on May 20, confirmed this trajectory: $81.6 billion in revenue, up 85% year over year, with guidance for the current quarter set at $91 billion.
The constraint side is a different story. The AI trade is running into bottlenecks - geopolitical, physical, financial - and none of them have an earnings call on their calendar.
The first and most immediate constraint is China.
A market that pays zero
China once accounted for at least one-fifth of Nvidia's data-center revenue. That share is now effectively zero, at least by the company's own accounting. Nvidia reported no China data-center compute revenue in its fiscal first quarter and is not assuming any in its current-quarter outlook either, with uncertainty over whether Beijing will allow any imports at all.
The Trump administration banned Nvidia H20 chip sales to China in April 2025, reversed course in July, and then last December approved the more advanced H200 for export to approved Chinese buyers, provided that the U.S. government received a 25% cut of sales. The Commerce Department published that framework in January 2026. The day after the rules appeared, Reuters reported that Chinese customs officers had been told not to let the chips in.
Roughly 10 major Chinese technology firms, including Alibaba $(BABA)$, Tencent (HK:700) and ByteDance, were cleared for purchases of up to 75,000 H200 units each. As of yet, Nvidia has not generated a single dollar of H200 revenue from China. Huang himself acknowledged that Nvidia now has what amounts to zero market share in China and said the export policy "has already largely backfired."
The structural damage to the addressable market
For Nvidia investors, the China situation is not simply a temporary revenue glitch. It is a structural change in how Nvidia's addressable market gets priced. Market share lost to Huawei's Ascend chips does not automatically return when licenses are granted, because Chinese hyperscalers are building procurement pipelines and software stacks around domestic alternatives. Every quarter of zero China revenue for Nvidia is a quarter in which Huawei's Ascend platform deepens its foothold - and the software ecosystems, training pipelines and procurement relationships built around domestic hardware do not dissolve whenever an export license finally arrives.
The H200 - a Hopper-generation chip that predates Nvidia's current Blackwell platform by a full architectural generation - is the most advanced hardware Washington has agreed to license for Chinese buyers. Nvidia's Blackwell platform and the forthcoming Vera Rubin remain off-limits for China. At Nvidia's current scale - $193.7 billion in data-center revenue in fiscal 2026 - one-fifth of that business represents a potential $38 billion annually sitting at zero. Huang himself has estimated the total Chinese AI chip market at $50 billion. China is the largest market Nvidia is currently locked out of, and the one now excluded from every forward estimate Nvidia itself has published.
The broader policy environment constitutes what one analysis described as a "permanent regulatory risk premium" on any semiconductor company with meaningful China exposure. Nvidia's market valuation assumes predictability. The Trump administration's China policy has delivered the opposite: H20 restrictions in April 2025, a reversal in July, H200 approval in December under new conditions, Chinese customs resistance and now congressional legislation seeking to override the executive order - all over just 13 months. No financial model handles that gracefully.
The buildout is running into its own walls
Data-center capacity is running up against power-grid limitations in nearly every major market.
The geopolitical bottleneck does not operate in isolation. It sits alongside a second layer of physical constraint that shapes the pace of the entire AI buildout. Data-center capacity is running up against power-grid limitations in nearly every major market. Grid interconnection queues in the United States currently run five years or more in most major markets. Cooling infrastructure - water rights, supply chains for chillers and heat exchangers - has become a project-level constraint. Advanced-node capacity at TSMC $(TSM)$ and its competitors remains tight through 2026 and into 2027, with AI accelerator demand now projected to consume the overwhelming majority of leading-edge wafer output through the end of the decade.
The scarcity argument that underpins Nvidia's pricing power is real, but scarcity in the physical infrastructure layer is not unambiguously bullish for investors. Every constrained watt and every delayed data-center rack is AI capacity that does not come online. It also means the AI buildout is competing for physical resources - land, energy, construction labor, raw materials - that are finite and already under pressure.
The financial bottleneck
At this scale, equity alone cannot fund the AI buildout. The bond market has been drafted.
The second structural constraint operates at a level most tech analysts do not model: the bond market.
The AI infrastructure buildout has crossed a threshold. The largest technology companies - the same hyperscalers whose capex guidance underpins Nvidia's forward revenue estimates - have shifted from returning capital to their shareholders to competing for it in the debt markets. Microsoft $(MSFT)$, Meta Platforms (META), Alphabet $(GOOG)$ $(GOOGL)$, Amazon.com (AMZN) and Oracle $(ORCL)$ have all issued long-duration bonds to fund data-center construction and GPU procurement at a scale and pace with no historical precedent in the technology sector. At this scale, equity alone cannot fund the buildout. The bond market has been drafted.
Tech-bond issuance competes directly with government debt and corporate debt from every other sector for the same pool of savings. The five major hyperscalers issued $121 billion in U.S. corporate bonds in 2025 alone - more than four times their prior annual average - and analysts at BofA project that figure to reach $175 billion in 2026, following Amazon's $54 billion global bond deal in March - the largest in the company's history. At roughly $700 billion in combined capex for 2026, these borrowing demands are large enough to move the market.
Every component of the AI buildout - power, silicon, steel, water - is in short supply and getting more expensive. More data-center construction means more energy demand, more materials, more labor. The bottleneck that inflates Nvidia's margins also inflates the cost of building the infrastructure that generates them. The AI trade's strongest fundamental argument - genuine scarcity - undermines the financial conditions required to sustain the trade's current valuation. For a stock at Nvidia's multiples, the further out the cash flows, the harder rates hit the valuation. Nvidia keeps beating estimates. The discount rate doesn't care.
What each bottleneck needs to clear
The premium that hyperscalers are paying to access capital is rising.
The geopolitical bottleneck has a clearance condition: a stable, predictable export framework that both the executive branch and Congress are willing to enforce consistently. Thirteen months of policy reversals on H20 and H200 have already cost Nvidia its entire China market share. What the market is waiting for is not simply a deal - it is a durable deal, the kind that survives an election cycle and does not dissolve the day after it is published.
The physical bottleneck clears only if power, grid, cooling and construction scale fast enough to keep the money being spent from sitting idle in a queue.
The financial bottleneck has a different clearance condition: bond markets that remain open and absorptive as hyperscaler debt issuance continues to scale. For now, investor appetite for tech paper is holding - Amazon's March deal was nearly four times oversubscribed. But the premium that hyperscalers are paying to access that capital is rising, and the bond market is beginning to price in the volume of supply coming behind it.
Huang moved the needle slightly on the first question last week. The China market will open, he said, "over time." Until these conditions clear, the AI trade carries risks that no quarterly earnings call can resolve. Huang can answer for the chips. But interest rates have no earnings call.
Also read: Nvidia delivers upbeat outlook, boosts dividend, but the stock falls. Here's what to know.
More: Artificial intelligence has Americans worried about jobs. Now there's a new AI worry.
-Jurica Dujmovic
This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.
(END) Dow Jones Newswires
May 21, 2026 11:05 ET (15:05 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments