AI News

After Jensen Huang hands over power, Nvidia's chain-leader credit and AI investment logic face revaluation

HuxiuUpdated Nov 16 21:02

AI summary

Light Cone Intelligence analysis holds that Jensen Huang's handover of power has changed the industry-chain credit role Nvidia has long played; supply and capacity commitments have risen from $119 billion to $279 billion, and the maximum exposure of AI cloud land, power and data-center guarantees has reached $108.5 billion. The article notes that Anthropic's revenue in the first nine months was $33.2 billion, with purchase and lease commitments of about $900.9 billion, while CoreWeave has a backlog of $132.5 billion yet a net loss of $709 million; data-center delays will flow directly into the income statement. The author argues that large tech companies and leading model companies can still adjust, while the new clouds and highly leveraged compute providers caught in the middle lack pricing power yet carry long-term debt and concentrated delivery obligations.

Why read this

AI score 76

The article separates Nvidia's guarantees from its payment-term arrangements, explaining how, after the chain leader's credit exits the stage, the time gap between AI orders and physical delivery is being repriced.

Topics

Tags

Body · AI translation

**Key Points**

**1. Jensen Huang's handover of power triggers a reassessment of the chain leader's credit**

NVIDIA's supply and capacity commitments rose from $119 billion to $279 billion, and its guarantees for AI cloud, power and other exposures reached $108.5 billion; his successor had pushed to lower the OpenAI backstop to about $105 billion, and the market is worried about the reliability of orders, payment terms and guarantees.

**2. AI investment credit is repeatedly pledged, and physical delivery becomes a hard constraint**

Anthropic's revenue in the first nine months was $33.2 billion, with purchase and lease commitments of about $900.9 billion; CoreWeave has a backlog of $132.5 billion but a net loss of $709 million and interest expense of $846 million. Data center delays will hit the income statement, and Anthropic also took over 1.7GW of gas equipment for about $3.4 billion.

**3. Alternative trades are unstable, and intermediate compute providers are under the most pressure**

AMD and Broadcom rose 18.93% and 18.16% respectively from November 6 to 13, while NVIDIA fell 12.75%, and then the two fell 8.75% and 4.76%. Large companies and leading model companies can adjust, while new clouds and highly leveraged compute providers lack pricing power but carry long-term debt. 2026-11-16 11:49 Jensen Huang wants to come back, but AI can't go back. Light Cone Intelligence quick overview

This article is from the WeChat public account: Light Cone Intelligence, author: a follower of frontier technology

We have simply been living in the previous era for too long

"Whatever." "I don't care."

At the final moment, NVIDIA CEO Jensen Huang was extremely exhausted. The man who had always been endlessly energetic, the man who signed for AI, the man who once personally intervened in supplier capacity, customer financing and policy, no longer cared at the final moment.

From the board's public investigation to the formal handover, it was only a few days. The company founded 33 years ago cut ties with Jensen Huang. This cut is also splitting the AI era in two.

Over the past half month, major AI news has come thick and fast. First Anthropic published its prospectus, and then its valuation was slashed by the market. Then came the U.S. midterm elections in early November, when voters fed up with AI data centers, together with capital, "smashed through" the market value of AI companies. Now NVIDIA, seen by the world as the foundation of AI compute, has also lost the captain who had been at the helm since day one.

AI large-model technology has not stalled, and enterprise customers have not canceled AI budgets overnight. IREN's latest results show AI cloud revenue growing sharply, with new contracts signed after the quarter, and prepayments on some new contracts covering up to about 90% of GPU capital expenditure; Google announced a long-term $17 billion investment plan in Wyoming. All of this points to demand still expanding.

But the industry's valuation standards have become chaotic. Wall Street used to price model parameters, GPU counts and cloud revenue. Now, electricity meters, transformers and ballots have joined the pricing team.

Technology continues to accelerate, but the certainty that organizes it is shrinking. Financial contracts can turn demand for the next decade into today's orders, and Jensen Huang's technology strategy and "face" can rally a large number of dream-filled scientists.

What Jensen Huang handed over

was not just a company

We can temporarily set aside GPUs and CUDA and understand NVIDIA from the perspective of AI deployment.

From the financial reports, it can be confirmed that NVIDIA holds considerable industrial power in the AI industry. As of July 26, 2026, NVIDIA's supply and capacity commitments rose from $119 billion in the previous quarter to $279 billion. For large data center purchases by some investment-grade customers, it can offer longer payment terms. NVIDIA also disclosed that guarantees for AI cloud land, power and computer rooms, as well as guarantees related to SB Energy, have a maximum total exposure of $108.5 billion.

What do these numbers mean? Take the SB Energy guarantee as an example. This guarantee takes effect gradually as projects are built in phases and leases begin, is triggered by specific default events, and covers only agreed lease and power payments. The point is that as long as NVIDIA is willing to put up its own credit, it can secure project implementation and exclusive adoption of NVIDIA architecture in the parks at very low risk.

The above case is just a lens reflecting NVIDIA's position in the AI industry chain. It sells GPUs to customers and also helps customers obtain the conditions to buy GPUs through investment, payment terms, cloud service procurement and guarantees. The identities of supplier, shareholder, customer and guarantor overlap in the same company.

For a rapidly expanding industry, NVIDIA's role is crucial. Without long-term commitments, suppliers dare not expand capacity, and data centers find it hard to finance. The market has almost converted these mutually supporting arrangements into certain growth.

A very important precondition for this system to work is Jensen Huang.

NVIDIA is an extremely flat organization, with no real "number two." Even more remarkable, in the 33 years since Jensen Huang founded and led the company, every technology strategy and every bet at a "life-or-death moment" for NVIDIA has been right. In the AI era, Jensen Huang explains the next generation of products to suppliers, proves to customers the necessity of compute investment, describes the AI vision to capital markets, and also fights in Washington for NVIDIA's qualification to continue selling.

Technology barriers make NVIDIA hard to replace, and Jensen Huang makes the entire industry chain believe that everyone is heading toward the same future.

Because many links do not require re-validating the entire AI industry; they only need to confirm that Jensen Huang is still standing at the center of the stage. The product roadmap he presents at launch events is often the supplier's production plan for the following year; the partnerships he announces together with customers are also seen by the market as a signal for the next round of capital expenditure. The more complex an industry chain is, the more it needs someone to align timelines on behalf of different parties.

But now, the person standing in this position has handed over power, and the object of trust in the industry chain has changed.

The previous orders are all still there, and the product roadmap will not change immediately either. Who guarantees the orders, where does the customers' money come from, how long are the payment terms, and under what circumstances can projects be renegotiated?

The same questions apply to the newly appointed Cress. As CFO, she had reservations about the large OpenAI backstop plan led by Jensen Huang, pushing to reduce the initially discussed scale of about $250 billion, with the final figure landing at about $105 billion. Jensen Huang called her a "bookkeeping accountant" in internal meetings, like "a deer frozen in the headlights."

Demand that is mainly sustained by the credit of the chain leader will gradually be priced separately from independent, verifiable end demand. The same applies to the supply chain. In the past, whoever appeared beside Jensen Huang and whoever was included in NVIDIA's product roadmap would affect a company's status.

Objectively speaking, this previous approach in the AI industry was not necessarily correct. It could improve efficiency, but it also concentrated a large amount of industry trust in one person. Having the evaluation of AI return to value is a good thing; after all, it is more robust and more logical.

There is just one problem: this shift in model arrived without warning.

After the news of Jensen Huang's resignation was announced, AMD and Broadcom rose for a time, and the market quickly told the story of "one whale falls, all things rise." Some analyses pointed out that if one looks only at the Philadelphia Semiconductor Index, the overall decline recently was not large, because rotation occurred within the index, and the market value evaporated from NVIDIA was offset by gains in AMD and Broadcom.

But at the level of absolute amounts, that is not the case.

As of November 13, NVIDIA's market value was about $3.92 trillion, down about $1.84 trillion from the October 6 close; from November 6 to 13, it fell about $573 billion. Over the same period, AMD and Broadcom had a combined market value of about $2.51 trillion, about 64% of NVIDIA's.

From November 6 to 13, NVIDIA fell 12.75%, while AMD and Broadcom rose 18.93% and 18.16%, respectively, and the Philadelphia Semiconductor Index rose 8.05%. Substitute trades did indeed appear. But by November 13, AMD and Broadcom had fallen 8.75% and 4.76%, respectively, showing that this rotation is not stable.

As overseas hardware stocks fluctuated again, investors also began to realize that substitutes can get orders, but may not be able to provide the same level of guarantees, payment terms, and industry confidence as NVIDIA.

From this perspective, Jensen Huang's handover of power has pushed this already-occurring structural adjustment to the forefront.

Last autumn, the hyperscale compute cooperation promoted by OpenAI and partners such as Oracle brought AI investment to another order of magnitude. At the time, the market generally cared about how large the orders were, how many GPUs there were, and which company got a ticket aboard. But the question truly worth caring about should have been: how many times have these revenues that have not yet occurred been used in advance?

In the earlier article "OpenAI Rallies the 'Trillion Club,' Opening the 2.0 Era of AI Investment," we mentioned that the key change in this event was that AI investment moved from the income statement to the balance sheet.

Model companies' expected revenue supports long-term compute procurement; compute procurement supports financing for new clouds and data centers; NVIDIA participates through orders, investment, payment terms, and guarantees; suppliers expand production accordingly; and capital markets then translate these plans into valuations.

In our interpretation of Anthropic's prospectus, we provided a clear cross-section of this chain. "Investors Who Cannot Afford to Lose Should Not Read Anthropic's Prospectus"

In the first nine months of 2026, Anthropic achieved revenue of $33.2 billion, up 1,206% year over year. At the same time, its total committed expenditures under purchase commitments and lease agreements were about $900.9 billion, almost all related to compute and infrastructure. Of that, about $451.6 billion is expected to occur before the end of 2030, and about $70.7 billion of commitments will mature within the 12 months after September 30, 2026.

Anthropic's revenue retains the elasticity of the technology industry, but its spending is no longer allowed to be easily changed.

The most eerie thing about this system of mutual trust is that orders, prepayments, capital expenditure, interest, and losses can all grow at the same time.

Take CoreWeave (a company whose main business is renting compute) as an example. In the third quarter, CoreWeave's revenue was $3.661 billion, and its revenue backlog was about $132.5 billion, but these backlogs still need to meet delivery and service availability requirements. The company also expects to receive about $9 billion in customer prepayments and deposits in the next three months. Over the same period, the company's net interest expense reached $846 million, its net loss was $709 million, and its 2026 capital expenditure guidance was as high as $37 billion to $39 billion.

The market used to be accustomed to treating orders as revenue, revenue as cash flow, and cash flow as debt-servicing capacity. But in the AI era, as long as every link continues to expand, the time gap in the conversion process will be temporarily covered by new contracts and financing, and then the left foot steps on the right foot and spirals into the sky.

But the future that capital can absorb ultimately has capacity.

Among these, "cloud computing" is probably the technology industry's most successful spatial rhetoric, making data centers sound like weightless assets.

Take Nebius as an example: the company's third-quarter revenue grew 571% year over year. But the shareholder letter also disclosed that some third-party projects have been delayed, and some 2026 revenue will shift to early 2027. One engineering delay thus entered the income statement.

This delay proves that the data center "GW" is a misleading number. Planned capacity, approved capacity, capacity under construction, grid-connected capacity, delivered capacity, and ultimately revenue-generating capacity are separated from one another by long implementation and delivery milestones.

In China, at this year's WAIC, super nodes, liquid cooling, and modular power supply and distribution offered some solutions. These engineering innovations can indeed increase power density, reduce energy consumption, and compress part of the deployment time. But what they solve is efficiency inside cabinets and campuses; they cannot take care of grid expansion and transmission lines.

Moreover, design changes and delivery delays in large projects are very common. But when equipment procurement has already been advanced, financing has already been secured, and stock valuations have long been lined up according to commissioning dates, if the physical world is a few weeks late, financial capital has to recalculate everything.

Looking further upstream, the pressure is likewise already "maxed out."

According to a Financial Times report, Anthropic took over about 1.7GW of gas-fired power generation equipment that OnePWR had previously ordered from Caterpillar for approximately $3.4 billion, and incorporated the engineering and construction capabilities needed for installation. A model company starting to directly purchase power generation equipment shows that physical constraints have begun to reshape model companies' asset structures.

This is almost like a paradox. Anthropic once expanded through talent, algorithms, and APIs; now, to obtain enough computing power, it also has to deal with generators, construction crews, and land contracts. The less grounded AI is, the more it must be grounded.

The physical world

ultimately has to pay the technical price

Large tech companies can build their own data centers, allocate capital across different regions, and have thicker cash flow buffers. Leading model companies can change suppliers, adjust procurement rhythms, and pass some pressure up and down the supply chain. Those suffering most are the new clouds, data center developers, and highly leveraged computing power operators caught in the middle. They cannot obtain the model companies' excess returns and lack pricing power over customers, yet they carry long-term debt, equipment orders, and concentrated delivery obligations.

Therefore, AI-related investment will become more rational going forward. Creditors demand more guarantees, customers demand stricter acceptance, and delays begin to affect asset valuations. In other words, the prospects of AI are judged separately from the physical world's capacity to bear them.

There are many specific projects involved, for example in land.

A working paper from the Dallas Fed this year estimated that by 2025, data centers had already raised average U.S. wholesale electricity prices by 6.9%, and in Northern Virginia the impact reached 15%. This increase is only an estimate of AI models' effect on the wholesale power market; the actual situation for residents still needs separate discussion.

Large models are good at handling billable tasks: writing code, optimizing products, solving math problems, and helping enterprises reduce labor costs. The efficiency gains they bring can quickly enter corporate financial statements. But unemployment transition, community impact, and infrastructure costs cannot all be handled by AI.

Moreover, the physical world is not only land and power grids. The phones, PCs, and AI glasses in users' hands require chips, batteries, optical components, manufacturing, and channels to update together. For models to enter enterprises, permissions, processes, and responsibility systems need to be rearranged. As AI begins to take on more cognitive tasks, job division and the labor market will also change accordingly.

Only when all of these are resolved will the AI era truly arrive. Just as no one questions power plants today, AI data centers should inherently be part of infrastructure. And this part may not achieve "big results" simply by attaching an AI label.

At present, many people compare this with the tech bubble of 2000. The internet back then also had revenue, and fiber optics did later carry the internet boom, but early builders and investors did not necessarily all live to see returns.

Conclusion

There is nothing new under the sun. More than 130 years is enough for humanity to take a similar script and act out the same fears, struggles, and expectations once again.

On the afternoon of October 11, 1889, a corpse on a wire triggered panic among ordinary Americans. The man was a repair worker; the wire he grabbed was supposed to be low-voltage, but several streets away it had crossed and short-circuited with a high-voltage line.

At that time, the electric age had just begun, and the wires over New York were as chaotic as the aftermath of this incident. The media wrote voluminously in newspapers that he was "first electrocuted, then burned." The mayor ordered a large number of high-voltage arc lamps turned off, and large areas of Manhattan returned to darkness. The capital camp led by Edison loudly proclaimed its case, and electricity was immediately drawn into the struggle over technical routes and commercial interests.

The changes of an era conceal enormous interests. At technological turning points, capital and market institutions will not let any opportunity slip by, trying to redistribute control.

Let alone ousting a CEO who holds only 3.58% of the shares. Even if the company he leads is the underlying foundation of the entire AI industry chain, even if his "face" is the adhesive that holds the AI investment alliance together.

Has published 184 articles on Huxiu. This content comes from the internet. Original link. The views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]

Source

Huxiu