On the afternoon of August 26, Silicon Valley time, NVIDIA released its financial results for the second quarter of fiscal year 2027.
First, the numbers: total revenue was $96.22 billion, up 106% year over year and 18% quarter over quarter. Wall Street had expected $92.17 billion, meaning NVIDIA beat expectations by more than $4 billion. Data center revenue reached $89.0 billion, down 117% year over year. Adjusted earnings per share were $2.22, versus expectations of $2.10. Net income more than doubled from the same period last year.
If you look only at these numbers, this was a flawless earnings report. But if all you see are the numbers, you are missing something more important.
What Does $89 Billion in Data Center Revenue Actually Mean?
$89 billion. In a single quarter.
What does that number mean? Put into a Chinese context, it is roughly equivalent to one-third of ByteDance’s total revenue for all of 2024. A chip company generated, in three months, revenue equal to one-third of what another company earned over an entire year — and what NVIDIA sells is still essentially the “raw material.”
More important is the composition. Of that $89 billion, hyperscalers contributed approximately $48.7 billion, accounting for nearly 55%. This means Microsoft, Google, Amazon, and Meta remain NVIDIA’s largest customers. Their combined capital expenditures in 2026 are expected to exceed $300 billion, and a substantial portion of that spending ultimately turns into NVIDIA revenue.
Interestingly, NVIDIA CEO Jensen Huang used one word during the earnings call: “incredible.” He described data center demand as “incredible.” He has used this word for at least four consecutive quarters. When a word is repeated so often that it loses its informational value, you instead have to ask: what exactly is it concealing?
Our observation is that it is not concealing insufficient demand — quite the opposite. What it conceals is that demand is overly concentrated.
Blackwell Ramps Up: The Generational Transition From Hopper to Blackwell
The biggest variable this quarter was the large-scale shipment of the Blackwell architecture.
The transition from Hopper (H100/H200) to Blackwell (B200/B300) is not simply a process-node upgrade. Blackwell’s core improvements lie in the fourth-generation iteration of the Transformer Engine, greater memory bandwidth, and more efficient inference optimization. In Jensen Huang’s own words, Blackwell further reduces the “inference cost per dollar.”
What does this mean for the developer ecosystem? Our assessment is that it is accelerating a trend: the economics of AI inference are moving from “usable” to “economically viable.”
Over the past two years, the developer community’s central complaint has been that training is too expensive, while inference is even more expensive. The ramp-up of Blackwell means that, at least for leading model providers and cloud companies, cost pressures at the inference layer are being systematically alleviated through hardware iteration. This is not a question of chip specifications. It is a question of business models.
If inference costs fall by another order of magnitude, a large number of AI use cases that are currently too expensive to run — from real-time translation to multimodal search, from code auditing to video understanding — will suddenly become viable. NVIDIA is not merely selling chips. It is selling the economic viability of the entire AI application layer.
China Goes to Zero: The Quietest Earthquake
Then there is the quietest, yet potentially most important, signal of the quarter: NVIDIA’s Q3 guidance includes no China data center compute revenue whatsoever.
Zero.
Q3 revenue guidance is $108.0 billion, with the midpoint representing growth of approximately 12% from the previous quarter. Wall Street had expected $104.2 billion. In other words, NVIDIA’s guidance for next quarter is nearly $4 billion above market expectations — and not a single cent of that $108 billion comes from China.
What does this mean? It means two things.
First, NVIDIA’s revenue growth has become completely decoupled from the Chinese market. Two years ago, China still accounted for a significant portion of NVIDIA’s data center revenue. Now that figure has fallen to zero, while total revenue continues to double. That in itself is an unsettling signal — it suggests that the structure of global demand for AI computing power is undergoing a fundamental geographic reconfiguration.
Second, it presents China’s AI industry with a very practical question: without NVIDIA chips, how should China’s AI infrastructure develop? Huawei Ascend, Cambricon, and a range of domestic alternatives have made significant progress over the past two years, but the generational gap with Blackwell in terms of performance density for high-end training clusters remains an objective reality.
The feedback we have heard from China’s developer community is this: in the short term, people are still relying on existing inventories of H100s and H200s to “stay alive”; in the medium term, domestic alternatives are moving from “usable” to “good to use”; but in the long term, if NVIDIA maintains the pace of iteration represented by Blackwell, the window for catching up is narrowing.
This is not a question that one earnings report can answer. It is an industry proposition that will require three to five years to validate.
The Subtle Signal in Gross Margin
Another easily overlooked figure: GAAP gross margin was 74.9%, while non-GAAP gross margin was 75.0%.
A 75% gross margin is almost counterintuitive for a chip company. The historical median for the semiconductor industry is roughly 45–55%. NVIDIA has been able to sustain this figure because it has virtually no competitor in the AI training chip market that poses a challenge in the true sense of the term — AMD’s MI300 series is catching up, but its market share remains limited; Google’s TPUs are used internally and do not constitute direct competition.
But there is a subtle signal here: gross margin did not continue to rise. In Blackwell’s first full quarter of large-scale shipments, gross margin was essentially flat. This could mean one of two things — either Blackwell’s early yields and production costs remain relatively high, or NVIDIA is actively managing pricing in order to preserve customer relationships and market share.
Whichever explanation is correct, 75% itself is already near the ceiling of the semiconductor industry. Our observation is that what the market truly cares about is not whether gross margin can go higher, but whether NVIDIA can hold it there.
Final Thoughts: The People Selling Shovels, and the People Mining for Gold
Back to the beginning.
$96.2 billion is an astonishing number. NVIDIA’s $108.0 billion Q3 guidance is even more astonishing. But more important than the numbers is the narrative: through one earnings report after another, NVIDIA is repeatedly proving one thing — AI is not a bubble. AI is an infrastructure revolution that is already underway.
But infrastructure revolutions have one characteristic: they do not distribute their victories evenly. The people selling the shovels (NVIDIA) making enormous profits does not mean that every miner (AI startup) will find gold.
From what we are observing, the AI industry in 2026 is entering a period of divergence: revenue at leading model companies is growing rapidly, while a large number of small and medium-sized startups are still struggling to achieve PMF (product-market fit). The better NVIDIA’s earnings become, the more visible this divergence becomes — because the high barrier created by compute costs is systematically determining who gets to remain at the table.
This is the real question that NVIDIA’s earnings report did not answer, but that everyone will have to confront.