#英伟达财报周 NVIDIA earnings +106%, stock up just 4%: What is the market afraid of?

NVIDIA's latest earnings report showed earnings up 106%, but its after-hours stock price rose only 4%.
The conclusion first: this contrast itself is the answer. It is not that the market does not recognize this report's performance; the performance had already been priced into the stock. What is truly being repriced is next year's figures, along with a business model shift that most people have overlooked.

I. The numbers themselves: Above expectations, but not above "expectations within expectations"
F2Q revenue was $96.2 billion, up 106% year over year, versus the market consensus of $94.6 billion—a beat, but not a large one. Data center revenue was $89 billion, gross margin was 75%, and EPS was $2.22; every metric beat expectations. These results would be explosive for any company, but for NVIDIA they merely count as "mission accomplished."
The logic is simple: Wall Street is an animal that buys expectations and sells reality. Over the past four quarters, every earnings report has beaten expectations, while the stock's reaction has diminished each time—the beat itself was priced in long ago, and even the earnings report became part of expectations. The real information lies in the guidance. F3Q guidance is $108 billion ±2%, versus market expectations of $105.1 billion. The moment management spoke, it was already $3 billion above consensus. This was not "we did a good job"; it was management telling the market in advance: Q3 will beat again. That 4% after-hours gain was mainly payment for that statement. But growth is shifting gears. Q3 growth is +12% quarter over quarter and approximately +78% year over year, compared with Q2's +106%; the curve has quietly turned downward. The story of consecutive doublings will begin to slow in Q3—this is the first data point in the entire AI narrative to start decelerating.
Gross margin is even more direct. Q3 guidance is 73.5%–74.5%, down from 75% in Q2, and is expected to fall further to 71%–72% in F4Q. Management did not sugarcoat it this time: HBM and memory costs are higher than expected, creating a structural headwind rather than a one-off disruption.
In other words, NVIDIA's profits are being transferred into the pockets of memory manufacturers.

II. The real bombshell: 70% in FY28
While everyone was focused on Q3's $108 billion, the CFO dropped something heavier on the conference call: FY28 revenue growth of 70% year over year. What is the market consensus? Around 45%. That is a 25-percentage-point jump in one stroke. This was not a beat; it was a change to "expectations" themselves.
That is the real reason for the 4% after-hours jump—not that this earnings report was particularly good, but that the market suddenly realized next year's growth curve is a full step higher than it had thought. The CFO added an even more forceful statement: 70% is the figure under supply constraints; if capacity keeps up, demand could support 100%. It was not saying "we will grow 70%"; it was saying "the market wants 100%, but we can only provide 70%." The bottleneck in AI has shifted entirely from the demand side to the supply side—silicon, packaging, HBM, and power. These four words are the treasure map for the global computing-power industry chain over the next two years. They also explain why inventory has reached a record $32 billion and receivables have been extended to 60 days.
Some people cite these two figures and say "demand is weakening," but they have it backward: NVIDIA is stockpiling HBM for Rubin's ramp-up and using looser payment terms to lock in orders from major customers. Demand has not weakened; NVIDIA is trading its balance sheet for supply.

III. Three things more valuable than the numbers
First, the business model. NeoClouds—third-party computing-power rental providers such as CoreWeave and Lambda—are shifting to revenue sharing: sell the hardware once, then share the rental revenue again, which management says will bring "tens of billions of dollars" in incremental revenue. The same chip generates revenue twice—from selling the shovel to selling the shovel while also collecting a toll.
This change is structural: NVIDIA's valuation model is beginning to include an additional layer of platform revenue.
Next, value per GW. In the Hopper era, revenue per GW was $18 billion; Blackwell is at $25 billion; Vera Rubin goes directly to $40 billion, more than doubling in two and a half years. When the revenue generated per megawatt doubles, the pricing anchors for every supporting link around computing power—PCBs, liquid cooling, power supplies, optical modules, and connectors—change completely. The sharp gains in A-share PCB, CPO, and liquid-cooling stocks today are not based on concept speculation; they reflect a revaluation of value per GW.
Finally, Rubin. It has already shipped. This is not a sample shipment or mass production next year; it is already running in customers' data centers. It is set to contribute 20% of data center revenue in F3Q, a full quarter earlier than market expectations, while also achieving the "fastest ramp-up in history." Blackwell's show is not over yet, and Rubin has already taken the stage—the industry's prosperity is not a matter of "future tense" but "present progressive."

IV. A-shares today: ¥2.13 trillion, competing for a face-up card
The ChiNext rose 1.71% to close at 3,473 points, the STAR 50 rose 3.77%, and trading volume across the two markets reached ¥2.13 trillion, an increase of ¥317.2 billion. The key point was the increase in volume. The previous two trading days saw low-volume rebounds—the indexes rose while turnover shrank, with existing funds playing among themselves. Today, incremental capital truly came in, and those who had stayed on the sidelines for a week began to believe in the rebound. Volume leads price; this signal matters more than the size of the gain. Capital flows were also clean: chips +¥49.6 billion, data centers +¥31 billion, and CPO +¥29.6 billion; the top five sectors by net inflows from major players were all AI computing-power hardware. Power-grid equipment, banks, and liquor fell. Power-grid equipment was hit by Trump's executive order banning foreign transformers from entering the U.S. power grid, while banks and liquor were merely along for the ride.
One sentence sums up today's market: AI dominated completely, with concentration in hot sectors extremely high. But one concern cannot be avoided: the gains were too rapid and too uniform. The ChiNext returned above its 20-day moving average, but overhead is filled with trapped positions, and some high-level stocks had already stalled in the afternoon. Of those scrambling to build positions today, at least half plan to sell into a higher open tomorrow. An oversold rebound and a trend reversal are two different things; so far, only the former has been confirmed.

V. Directions worth watching next (no individual stocks)
Ranked by certainty and speed of realization, here are several directions.
The strongest is the entire AI computing-power hardware chain.
Rubin's $40 billion per GW is 2.2 times Hopper's, and this revaluation will continue through FY28. High-layer-count PCBs, high-speed connectors, liquid cooling, and high-voltage power supplies will all benefit equally from the revaluation of value per GW, with earnings realization coming from the end of this year into next year.
Next is domestic HBM and high-end memory.
NVIDIA itself acknowledged that HBM costs were higher than expected and had compressed gross margins. Globally, only SK hynix, Micron, and Samsung can supply it. The top-tier customers are competing for capacity and are insensitive to price, while China's domestic substitution rate is nearly zero—self-reliance and controllability combined with upstream scarcity make this the strongest domestic-substitution direction over the next 12 to 24 months.
Third, domestic computing-power chips.
There was an inconspicuous but very concrete figure in the earnings report: data center revenue from China accounted for less than 1%, and the Q3 guidance directly excludes China. NVDA has essentially been reduced to zero in China. For domestic chipmakers, this is not "potential substitution"; there is "no other choice." In addition, Vera CPUs are already in mass production, and CPU revenue will double in FY28—the validation of full-stack solutions as customers' ultimate need will raise the valuation anchor for manufacturers pursuing a full-stack approach.
Fourth, computing-power rental and third-party IDC.
NVIDIA is supporting NeoClouds worldwide and using revenue sharing to bind the entire ecosystem. In A-shares, this maps to computing-power rental and AIDC, directly comparable to CoreWeave and Lambda. China is still at the storytelling stage; once a genuine company delivers actual earnings, the room for repricing will open up.
Finally, CPO.
In the Rubin era, clusters will grow from tens of thousands of GPUs to hundreds of thousands, causing an exponential explosion in interconnect demand. Networking's 18% quarter-over-quarter growth has already provided validation. But CPO rose the most sharply today, and short-term positive news has largely been priced in—the direction is right, but the timing needs to be watched.

VI. Four conclusions, for the record
One, today's A-share market is an oversold rebound, not a trend reversal.
Two, the long tail of NVIDIA's earnings report lies in China. If FY28 really reaches 70%, global computing power will split between the "haves and have-nots," while China is the world's second-largest computing-power market. The gap will not disappear; it will only be filled in another form—the domestic-substitution narrative is just beginning its main upward wave.
Three, HBM is the hidden thread. Record-high inventory, extended payment terms, and compressed gross margins all point to the same explanation: NVIDIA is using every means available to lock in HBM capacity. This is a certain positive for global HBM manufacturers and an opportunity for domestic memory to "follow the pricing."
Four, stop asking whether NVIDIA is a bubble. Earnings are still accelerating at +106%, next year's growth is expected to reach 70%, and demand is at 100% of supply—this is a seller's market. A bubble is when prices become detached from fundamentals; now the fundamentals are chasing the price. The direction is reversed. $NVDA ‌