The amazing NVIDIA. No matter how good its financial report is, its stock price doesn't rise. No matter how fierce the short - selling is, its stock price doesn't fall.
NVIDIA's financial report for the first quarter of fiscal year 2027 is out. As expected, it still amazes me.
The revenue reached $81.6 billion, a year-on-year increase of 85% and a quarter-on-quarter increase of 21%. The GAAP net profit was $58.3 billion, a year-on-year increase of 211%; the non-GAAP net profit was $45.5 billion, a year-on-year increase of 139%. The gross profit margin was 74.9%, remaining stable between 74% and 75%.
The data center business generated $75.2 billion in revenue, accounting for 92% of the total revenue and a year-on-year increase of 92%. The quarterly dividend soared from $0.01 to $0.25, a 2400% increase. An additional $80 billion in stock buyback authorization was added.
Jensen Huang said in the earnings conference call, "The construction of AI factories is accelerating at an astonishing speed. This is the largest infrastructure expansion in human history."
However, NVIDIA's stock price fell by 0.77% after the release of the financial report.
This has been the case for three consecutive quarters. Clearly, NVIDIA's revenue exceeded the midpoint of the company's guidance by about $3.6 billion, but the stock price fell after the earnings report.
But then again, 99% of those who tried to short NVIDIA have suffered heavy losses.
Michael Burry, the prototype of the big short in the movie, disclosed his put option positions on NVIDIA and Palantir in the third quarter of 2025, with a notional value of about $1.1 billion according to the 13F filing.
He holds 1 million put options on NVIDIA with a strike price of $110 and an expiration date in 2027. As of May 8, 2026, NVIDIA's stock price was about $215, significantly higher than the strike price range of the relevant put options.
Leopold Aschenbrenner, a former OpenAI researcher, was even more aggressive in the first quarter of 2026.
His hedge fund, Situational Awareness, disclosed put option positions on NVIDIA, Broadcom, etc. With a notional value of billions of dollars according to the 13F filing, the notional value of the put options related to NVIDIA was about $1.6 billion.
As a result, NVIDIA's stock price rose by about 15% compared to the end of the first quarter, and Leopold got trapped.
The stock price doesn't rise even with excellent earnings reports, and it doesn't fall even with aggressive short - selling. Isn't it amazing?
The data center business is still booming
The most prominent figure in NVIDIA's financial report is the $75.2 billion revenue from the data center business, a year-on-year increase of 92% and a quarter-on-quarter increase of 21%.
Breaking it down, the revenue from data center computing was $60.4 billion, a year-on-year increase of 77%. The revenue from data center networking was $14.8 billion, a year-on-year surge of 199% and a quarter-on-quarter increase of 35%.
The data center networking business is experiencing explosive growth.
The reason lies in the products. Take the upcoming mass - produced Vera Rubin system as an example. A single system contains 1.3 million components, 72 Rubin GPUs and 36 Vera CPUs.
For so many chips to work together, a large number of high - speed network devices are needed to connect them. As the system scale expands, the demand for network devices also increases.
The Vera Rubin is an AI supercomputer system launched by NVIDIA in 2026 for training large - scale AI models.
Jensen Huang specifically mentioned in the earnings conference call that the Vera Rubin "has a very good start" and is expected to be more successful than the Grace Blackwell.
"Our market share in the inference market is growing very fast, and the number of cutting - edge model companies is increasing." At the same time, Jensen Huang also named Anthropic, saying it is NVIDIA's key new customer this year, providing models for Microsoft Azure, Amazon AWS and CoreWeave.
Jensen Huang also said that he expects the Vera Rubin "to be in a state of supply limitation throughout its entire life cycle."
The Grace Blackwell has long been sold out, and the Vera Rubin has been fully booked even before large - scale shipments. This is where NVIDIA's main growth engine comes from.
In sharp contrast to the data center business is the gaming business.
Previously, NVIDIA was a complete gaming company. Almost whenever there was something related to gaming, you could see NVIDIA's presence. But now, NVIDIA has little to do with gaming. In fact, it can be said that gaming should stay away from NVIDIA.
In fiscal year 2020, the gaming business accounted for more than 50% of NVIDIA's revenue, while the data center accounted for only 27%. In just a few years, this ratio has completely reversed. Now, the gaming business accounts for less than 8% of NVIDIA's total revenue.
In addition, NVIDIA's edge computing division generated $6.4 billion in revenue, a year-on-year increase of 29% and a quarter-on-quarter increase of 10%. This division includes PCs, game consoles, workstations, AI - RAN base stations, robots and cars.
NVIDIA urgently needs HBM: From HBM3E to HBM4
But now Jensen Huang has a problem that he can't solve, and that is HBM.
The Blackwell requires HBM3E, and the next - generation Rubin requires HBM4.
This is a type of memory made by stacking a large number of DRAM chips, which can achieve fast temporary data storage and allow chips to run multiple tasks in parallel. It is a key component of NVIDIA's GPUs.
The problem is that almost all of the world's HBM is in the hands of SK Hynix, Samsung and Micron.
Moreover, HBM cannot be manufactured on standard DRAM memory production lines. It requires specialized equipment, different processes and independent production capacity.
Kim Ki - tae, the head of HBM sales and marketing at SK Hynix, said, "The HBM demand in the next three years far exceeds our supply capacity." Sanjay Mehrotra, the CEO of Micron, said, "Our HBM production capacity for 2025 and 2026 has been fully booked."
In theory, the supply of HBM3 increases by 50% to 60% every year. However, the problem is that due to the increasing demand for NVIDIA's GPUs, the demand for HBM3 increases by 80% to 100% every year.
The demand far exceeds the supply, and the gap is getting bigger and bigger.
SK Hynix is investing more than $50 billion to build new HBM production capacity, Samsung is investing $40 billion, and Micron is investing $33 billion. But it takes 18 to 24 months for a new semiconductor factory to start production after construction begins.
Moreover, for an HBM factory to be profitable, its yield rate needs to reach over 85%. But this yield rate doesn't reach 85% overnight; it needs to increase gradually.
In the first half - year, it is called the "trial production stage", and the yield rate will increase from 30% to 60% - 70%. Around one year is called the "early mass - production stage", and the yield rate can only reach 80% at this time.
Then it takes about another year to reach over 90%.
That is to say, NVIDIA still needs a long time to alleviate the problem of HBM3 shortage.
However, in addition to the production capacity problem, there are also some social problems.
In May 2026, Samsung Electronics faced the largest - scale strike crisis in its history. 93.1% of more than 66,000 union members voted in favor of the strike. The union demanded the removal of the performance bonus cap because Samsung's competitor, SK Hynix, had removed the bonus limit, and some employees' bonuses were more than three times that of Samsung employees.
On the morning of May 20, the labor - management negotiation broke down again. The union announced a full - scale strike from May 21 to June 7, lasting for 18 days. The Bank of Korea warned that if the strike continued, South Korea's economic growth this year could be dragged down by 0.5%. The South Korean government estimated that one day of shutdown would cause a loss of about 1 trillion won.
But on the evening of May 20, in the labor - management negotiation that restarted in the afternoon of the same day, the two sides reached a tentative agreement. The large - scale strike originally scheduled to start on May 21 was postponed. The union will conduct an internal vote on the plan, and subsequent actions will be decided based on the vote result.
South Korean institutions once speculated that if the Samsung strike was cancelled and the labor - management relationship eased, the strike would cause the price of HBM3 to increase by about 20% - 30%; if the strike restarted but was restricted by the court, with partial production capacity fluctuations, the price would increase by 40% - 50%; if there was a full - scale strike and the Pyeongtaek campus shut down, the price would increase by 80% - 100%.
In conclusion, these remaining labor problems will inevitably be reflected in the price of HBM3.
In 2024, the price of a 36GB HBM3E was about $500. Today, the price of the same product has reached $1200, a 130% increase.
Not only that, in Q1 2026, the contract price of HBM increased by 25% - 30% quarter - on - quarter, and it is expected to increase by another 30% - 50% in Q2.
SK Hynix, Samsung and Micron control the throat of the supply chain. Since only these three companies have HBM3, NVIDIA's pricing power is now not in Jensen Huang's hands but in the hands of the memory manufacturers.
As the proportion of the data center business in NVIDIA's revenue increases, HBM3 will also directly determine NVIDIA's fate.
OpenAI and Anthropic are the real big customers
There is also a very dangerous figure in NVIDIA's financial report.
Colette Kress, NVIDIA's CFO, said that hyperscale cloud service providers account for 50% of the data center business.
Although not named, we all know in our hearts that they are Microsoft, Amazon, Google and Oracle.
This means that if any one of them cuts AI capital expenditure, NVIDIA's growth curve will be immediately impacted.
Fortunately, the four companies, Google, Amazon, Meta and Microsoft, are expected to spend $725 billion on AI - related capital expenditure in 2026, and this figure may exceed $1 trillion in 2027.
However, if you look closely at the businesses of these four companies, you will find such a fact. They are not four companies; in fact, they are two companies, OpenAI and Anthropic.
Microsoft is the largest cloud service provider for OpenAI. Almost all of OpenAI's GPT model training and inference run on Microsoft Azure.
Although Amazon AWS and Google both have their own self - developed chips, in order to meet Anthropic's computing power requirements, they will adopt a hybrid solution, with part of the computing power coming from the NVIDIA GPUs they purchased.
Oracle signed one of the largest cloud computing contracts in history with OpenAI, worth $30 billion, for a period of about 5 years.
Oracle is also one of the first customers to receive NVIDIA's Vera CPU. It plans to deploy hundreds of thousands of Vera CPUs starting in 2026 and is the first cloud service provider to commit to large - scale deployment of Vera.
Even Elon Musk is now a service provider for Anthropic. On May 6, 2026, Anthropic reached an agreement with SpaceX to rent all the computing power of SpaceX's Colossus 1 data center in Memphis, Tennessee, including 220,000 NVIDIA GPUs (H100, H200 and GB200), with a total computing power of 300 megawatts.
Therefore, there is a very delicate situation here. Both OpenAI and Anthropic are trying to get rid of NVIDIA.
In January 2026, OpenAI signed a $10 - billion agreement with the AI inference chip supplier Cerebras to obtain 750 megawatts of computing capacity.
Just four months later, in May 2026, OpenAI made an additional investment and signed a three - year agreement worth more than $20 billion with Cerebras. As part of the deal, OpenAI will acquire about 10% - 11% of Cerebras' equity and upgrade the supplier relationship to a strategic alliance.
Cerebras completed its IPO on May 14, 2026, with a market value of nearly $100 billion on the first day.
Now, the relationship between OpenAI and Cerebras is deeply bound. Sam Altman is an early investor in Cerebras and holds millions of dollars' worth of shares.
The situation with Anthropic is even more obvious.
In April 2026, Amazon invested $33 billion in Anthropic. Anthropic promised to spend more than $100 billion on AWS infrastructure in the next decade and will obtain up to 5 gigawatts of computing capacity, covering Amazon's Graviton CPU and Trainium2 to Trainium4 AI chips.
Google also promised to provide Anthropic with $40 billion worth of computing resources at the same time, including 3.5 gigawatts of next - generation TPU production capacity.
What's more ironic is that NVIDIA invested $30 billion in OpenAI and $10 billion in Anthropic in 2026. Logically, these two companies should be NVIDIA's "little brothers".
But in reality, NVIDIA's big customers now have to follow the computing power requirements of these two "little brothers".
Not only that, these two "little brothers" are turning the tables.
On May 5, 2026, OpenAI jointly released the MRC (Multipath Reliable Connection) protocol with NVIDIA, AMD, Broadcom, Intel and Microsoft. This is an Ethernet transmission protocol designed specifically for AI data centers.
OpenAI is no longer just a chip buyer; it has become the leader. Chip giants like NVIDIA, Intel and Broadcom now have to design products according to OpenAI's technical standards.
Anthropic is even more direct.
It not only uses Amazon's Trainium chips but has also deeply participated in the design of Trainium.
Anthropic works closely with Amazon's Annapurna Labs (Amazon's self - developed chip department) and directly provides feedback data on Claude training loads, thereby changing the architectural decisions of Trainium3 and Trainium4.
Amazon's engineering team and Anthropic communicate almost every day. From bottom - level optimization to high - level architectural decisions, Anthropic has a say.
In other words, the two companies invested in by NVIDIA, one is defining the network standards for AI data centers, and the other is participating in the design of competitors' chips. They have gained the right to speak.
Of course, NVIDIA is still the infrastructure of the entire AI era. NVIDIA's GPUs are still the gold standard for training large - scale language models, and the lock - in effect of the CUDA ecosystem is still strong.
