Next fiscal year, Nvidia must finance a quarter of its own revenue
Nvidia packages GPUs as financeable assets while funding AI labs and extending credit to customers buying its chips, propping up revenue through circular financing and driving accounts receivable to $63.1 billion.
The argument · tap a timestamp to hear it
Computing power is revenue—Jevons Paradox means cheaper chips drive volume growth
Huang has inverted the old accounting. GPUs were once an expense to minimize; now they're production assets that generate returns. But his wager rests on Jevons Paradox: as compute gets cheaper, the surge in demand activates new use cases—contract reviews, email, simulation—that were economically unviable before. Volume must grow fast enough to offset falling prices. If it does, Nvidia widens margins despite cheaper chips.
— Lao ShiAI agents never sleep—they parallelize and multiply compute demand by orders of magnitude
Agents differ from humans: they reason, plan, call tools, iterate repeatedly, work without sleeping or procrastinating, and can spawn hundreds of copies running in parallel. An engineer might manually run dozens of simulations a day; an agent can generate thousands of parameter combinations and optimize them autonomously. Human software interactions are dropping while background compute demand explodes. Agents don't just need GPU—they need CPU for scheduling, memory and storage for state, networking to connect tools and systems. One machine's gains lift everything around it.
— Lao ShiTokens are intermediate products; value is solving problems correctly, not token count
A steel mill produces a ton of steel, sells it, makes cash. An AI factory produces ten million tokens that might be worth nothing. An agent might spin for an hour burning millions of tokens to fix a bug, only to make the code worse, forcing engineers to spend extra time cleaning up the mess. Nvidia cares about tokens-per-watt and tokens-per-dollar. Users care about spending money to solve one thing correctly. When they align, everyone wins. When algorithm innovation lets models use fewer tokens for the same task, that's good for users—but not necessarily for vendors selling hardware by token throughput.
— Lao ShiPer gigawatt, Nvidia captures revenue from $18 billion to $40 billion across chip generations
The AI factory's total efficiency—not a single GPU—determines pricing power. By Nvidia's own math, for each gigawatt of power a customer receives, the device-purchase revenue flowing to Nvidia was approximately $18 billion in the Hopper era, $25 billion in Blackwell, and $40 billion in the Vera Rubin generation. That extra money isn't higher GPU prices. It's that Nvidia has more to sell: networking, software, the full stack ecosystem. Even if customers use competitors' GPUs or custom silicon, networking and systems still flow through Nvidia. It keeps collecting rent.
— Lao ShiDSXPORT cuts data center deployment from six months to as little as two weeks
DSXPORT pre-assembles entire AI factories in modular containers—racks, wiring, liquid cooling, system validation—all done in the factory. At the site, power, cooling, and networking just need to be ready. Deployment time drops from six months to as little as fourteen days. Huang's calculation: an AI factory generates $40 billion annually in revenue; three months of delay is tens of billions in sunk loss. But gross margins on infrastructure components don't match GPU margins, and this quarter's 75% overall gross margin may face seasonal pressure ahead.
— Lao ShiCircular financing inflates revenue while extending credit disguised as long payment terms
Nvidia's management keeps repeating three words—productive, fungible, durable—to convince Wall Street that GPUs are stable, long-term, resellable assets. Customers can then use equipment as loan collateral. Nvidia itself invests heavily: supply and capacity commitments now exceed $279 billion, plus nearly $50 billion into AI labs. This quarter's $96.2 billion revenue grew 106% year-over-year. Yet accounts receivable hit $63.1 billion with collection periods around sixty days. More revenue isn't real cash. It's extended as long payment terms, accounting entries sitting with customers. Nvidia is using its own balance sheet to funnel hidden loans.
— Lao ShiEvery AI strategy—U.S., China, Europe—must pay Nvidia's infrastructure toll regardless of approach
The U.S. pursues frontier models plus high-value services; the challenge is whether billion-dollar investments pay back and custom chips reduce Nvidia reliance. China pursues open ecosystems and extreme engineering efficiency; DeepSeek and other open models lower the barrier to entry, paired with manufacturing-sector AI deployment; the challenge is competing under compute constraints. Europe pursues industrial applications and sovereign AI—adapting open models to local industrial data for domain-specific training. The earnings report signals the outcome: ACIE customers (non-hyperscaler data centers) contributed roughly $40 billion, up 138% year-over-year, outpacing overall data-center growth. Every path requires paying Nvidia's toll.
— Lao ShiNvidia's model survives only if three conditions hold: productivity gains, capital access, market share
Three balance sheets determine how long Huang's economics last. The technology account asks: does AI actually raise productivity, or does it stop at keynote demos? The capital account asks: is AI infrastructure expansion funded by customer subscriptions or increasingly by Wall Street financing and Nvidia's own balance sheet propping it up? Circular financing that re-invests its own returns is a snowball that eventually needs an accounting. The industry account asks: after markets mature, will open models and application companies capture a larger slice? Huang's historical reminder: the biggest winner in a technological revolution isn't always whoever solved the infrastructure bottleneck first.
— Lao ShiIn their own words · checked verbatim
The rules of the AI era have changed. Computing power is no longer a consumable; it's a productive asset that directly makes money for you.
AI时代的规矩变了 AI算力不再是消耗品 而是能够直接帮你赚钱的生产资料
Lao Shi1:02
As compute gets cheaper, it's not just that existing work costs less—it's that a whole class of tasks that were previously uneconomical suddenly become viable.
算力便宜了 不光是原来的活干的更便宜 而且一大堆原本用AI来做不划算的活 突然就变得划算了
Lao Shi3:02
What's more, agents don't run on GPU alone. They need CPU for task scheduling, memory and storage for data, and networking to connect tools and systems. One machine's gain lifts everything around it.
而且呢智能体干活 不是只靠GPU 它还需要靠CPU 去调度任务 需要内存和存储来存数据 需要网络来连接 各种工具和系统 这个就叫 一人得到鸡犬升天
Lao Shi4:02
According to Nvidia, if a site has power, cooling, and networking ready, on-site deployment time can shrink from six months to as little as fourteen days.
按照英伟达的说法 只要场地的电 冷却 还有网络都准备好了 现场部署的时间 就能从原来的6个月 缩短到最快14天
Lao Shi8:03
In the latest earnings call, Nvidia management repeatedly emphasized three words—especially productive and fungible, plus durable. They mean Nvidia's platform is universal enough to do training, inference, and run agents.
英伟达管理层在最新的财报会议上 反复在念叨三个词 特别是productive fungible 还有durable 他们的意思是 英伟达的平台足够通用 能做训练 能做推理 还能跑智能体
Lao Shi11:06
In plain terms: more and more of Nvidia's revenue never becomes actual cash. Instead it sits with customers as long payment terms. Nvidia is essentially using its own balance sheet to extend hidden loans to customers.
那把前面翻译成大白话 就是英伟达越来越多的收入 其实并没有变成真金白银 而是以长涨期的形式 压在客户那 英伟达相当于 在用自己的家底 给客户提供一种隐形的贷款
Lao Shi14:08
But 200 years of industrial history shows repeatedly that the value created by a foundational technology eventually disperses across all industries using it. The biggest winner in a technological revolution isn't necessarily whoever first solves the infrastructure bottleneck.
但是过去200年的工业史一次又一次的告诉我们 一项基础技术创造的价值 最终都会分散到无数使用它的行业 技术革命最大的赢家 不一定是最早解决基础设施瓶颈的那家公司
Lao Shi20:14
Figures
| Hopper era, per GW device purchasing | approximately $18 billion | 7:03 |
| Blackwell era, per GW device purchasing | approximately $25 billion | 7:03 |
| Vera Rubin era, per GW device purchasing | approximately $40 billion | 7:03 |
| DSXPORT on-site deployment time | reduced from 6 months to as little as 14 days | 8:03 |
| This quarter gross margin | 75% | 9:03 |
| This quarter revenue | $96.2 billion, year-over-year growth 106% | 13:08 |
| Accounts receivable | $63.1 billion, average collection period approximately 60 days | 13:08 |
| Next fiscal year, AI labs' share of revenue contribution | approximately one quarter | 14:08 |
| ACIE customers' revenue | approximately $40 billion, year-over-year growth 138% | 17:10 |
Glossary
- Jevons Paradox
- When technology improves efficiency and costs fall, total consumption often rises as new use cases become economically viable, paradoxically increasing overall usage.
- DSXPORT
- Nvidia's modular, pre-assembled data center product that ships as containerized infrastructure units, slashing deployment time from months to weeks.
- Productive / Fungible / Durable
- Three properties Nvidia emphasizes to convince financial institutions that GPUs are long-term, tradeable assets suitable as collateral for customer loans.
- circular financing
- Nvidia invests in or funds customers, who then use that capital to purchase Nvidia equipment, creating a self-reinforcing financial loop.
- ACIE
- Nvidia's category for data center customers beyond large cloud providers: AI labs, industrial customers, and enterprises requiring customized infrastructure.
How to listen
Investors concerned about AI infrastructure bubbles and Nvidia's financial health, plus founders and engineers wanting to understand compute pricing dynamics.
The closing Q&A and calls to action contain no new information.