May 24, 2023. The earnings release NVIDIA put out for the first quarter of fiscal 2024 (February–April 2023) carried one particular number.
Revenue for the quarter was $7.192 billion, of which the Data Center business accounted for $4.28 billion. What caught the eye was the "Outlook" section below it.
"Revenue is expected to be $11.00 billion, plus or minus 2%."
It amounted to advance notice that revenue would grow by nearly $4 billion in a single quarter. In the same release, Jensen Huang said, "A trillion dollars of installed global data center infrastructure will transition from general purpose to accelerated computing as companies race to apply generative AI into every product, service and business process." He also said, "We are significantly increasing our supply to meet surging demand."
Three months earlier, in the February 2023 earnings release, Huang had already said that "AI is at an inflection point." On March 21, 2023, announcing the cloud service "DGX Cloud," he said, "We are at the iPhone moment of AI." Further back, in March 2022, announcing the new GPU "H100," he had put it this way:
"Data centers are becoming AI factories—processing and refining mountains of data to produce intelligence."
The actual result went beyond even the $11.0 billion outlook. The second-quarter results announced three months later, on August 23, 2023, showed revenue of $13.51 billion, with the Data Center business alone at $10.32 billion (up 141% from the previous quarter and up 171% from a year earlier).
Why did companies keep buying this company's chips on such a scale? The clues lie in three things that had been piling up long before the generative AI wave arrived: the chips themselves, the network that links the chips together, and the software layered on top.
A sprint to a "one-year rhythm," and Mellanox as the blood vessels
The first was reshaping the GPU around the computations of deep learning.
On May 10, 2017, NVIDIA announced the "V100," a GPU on its "Volta" architecture. Packing 21 billion transistors, the chip carried 640 "Tensor Cores," a new unit created to speed up AI calculations, and claimed 120 teraflops of deep learning performance.
On May 14, 2020, the company announced the "A100," with more than 54 billion transistors, saying it was already in full production and shipping to customers worldwide, and explained that it raised performance by up to 20 times over the previous generation (according to the company's announcement). On March 22, 2022, it announced the "H100," with 80 billion transistors. The H100 carried a "Transformer Engine" that ran the calculations of the deep learning model known as the "Transformer" up to six times faster than the previous generation.
On March 18, 2024, it announced "Blackwell," with 208 billion transistors, along with the "GB200 NVL72," which put 72 Blackwell GPUs and 36 of the company's own "Grace" CPUs into a single rack, and said it delivered up to 30 times the performance of the same number of H100s for large language model (LLM) inference.
On June 2 of the same year, in a keynote ahead of the Computex trade show in Taipei, Huang said: "Our company has a one-year rhythm."
After that, in March 2025, NVIDIA announced "Blackwell Ultra," saying it would be available through partners from the second half of that year. At CES on January 5, 2026, it announced that the next generation, "Rubin," was "in full production." Yet the fiscal 2026 10-K, filed the following February, said Rubin "is expected to commence production shipments in the second half of fiscal year 2027," and the 10-Q in August said, "Our next-generation Data Center architecture, Vera Rubin, began production shipments in the third quarter of fiscal year 2027." The company's filings described the same product in different terms at different times.
The second was the network linking the chips together.
In March 2014, NVIDIA had announced plans to build into future GPUs "NVLink," an interconnect that would make data exchange between GPU and CPU 5 to 12 times faster than was then possible. In March 2018, it announced "NVSwitch," a switch that let up to 16 V100s communicate simultaneously at 2.4 terabytes per second, and "DGX-2," a system that used it.
Then, on March 11, 2019, NVIDIA announced an acquisition: an agreement to buy Mellanox Technologies, maker of products for "InfiniBand," the communications standard used in supercomputers, and for high-speed Ethernet, for $125 per share in cash, an enterprise value of about $6.9 billion.
The joint announcement sets out the picture of the data center the two companies had in mind.
"Datacenters in the future will be architected as giant compute engines with tens of thousands of compute nodes, designed holistically with their interconnects for optimal performance."
It meant owning not just the GPUs, the brains, but also the blood vessels that connect them. After the expiry of the U.S. antitrust waiting period, approvals from the European Commission and Mexico, and a conditional approval from China's State Administration for Market Regulation (SAMR), the acquisition closed on April 27, 2020, at about $7 billion. SAMR, finding that the combination had or might have the effect of eliminating or restricting competition in markets including GPU accelerators and specialized network equipment, had attached conditions such as no tying of products and continued supply to China. This notice comes up again in Part 5.
The weight of networking shows in the revenue tables. According to the fiscal 2026 10-K, "Networking" within Data Center revenue was $8.575 billion in fiscal 2024, $12.99 billion in fiscal 2025 and $31.376 billion in fiscal 2026. In the first quarter of fiscal 2027 (February–April 2026), it reached $14.8 billion in a single quarter (up 199% from a year earlier).
Not every acquisition bore fruit.
On September 13, 2020, less than five months after closing the Mellanox deal, NVIDIA announced that it would buy Arm, the British chip-design company, from SoftBank Group for a total of $40 billion in stock and cash.
This time, regulators in several jurisdictions stood in the way. On April 19, 2021, the responsible British minister issued an intervention notice on national-security grounds, and on October 27 that year the European Commission opened an in-depth investigation. Margrethe Vestager, the executive vice-president in charge of competition policy, said, "Whilst Arm and NVIDIA do not directly compete, Arm's IP is an important input in products competing with those of NVIDIA." On December 2, 2021, the U.S. Federal Trade Commission (FTC) sued to block the deal.
On February 8, 2022, NVIDIA disclosed that it had terminated the Arm purchase agreement because of "significant regulatory challenges preventing the consummation of the transaction" (8-K). The $1.25 billion it had prepaid would stay with the sellers, and NVIDIA wrote that it intended to record a $1.36 billion charge. What it kept was a 20-year Arm license.
From $1 trillion to $5 trillion: the financial statements rewritten
The generative AI wave that began in 2023 went on to rewrite NVIDIA's financial statements on a grand scale.
The first thing that happened was a change of lead. Through fiscal 2022 (the year ended January 2022), the Gaming business ($12.462 billion) was larger than the Data Center business ($10.613 billion). In fiscal 2023 the two traded places (Data Center $15.005 billion, Gaming $9.067 billion), and in fiscal 2024 Data Center reached $47.525 billion, more than three times the year before, pulling far ahead of Gaming ($10.447 billion).
From there, the figures year by year were as follows.
- Fiscal 2024 (year ended January 2024): revenue $60.922 billion (Data Center $47.525 billion) / gross margin 72.7%
- Fiscal 2025 (year ended January 2025): revenue $130.497 billion (Data Center $115.186 billion) / gross margin 75.0%
- Fiscal 2026 (year ended January 2026): revenue $215.938 billion (Data Center $193.737 billion) / gross margin 71.1% (a $4.5 billion charge related to the restrictions on H20 sales to China was recorded in the first quarter of that year)
Into fiscal 2027, revenue was $81.615 billion in the first quarter (February–April 2026; Data Center $75.2 billion) and $96.221 billion in the second quarter (May–July; Data Center $89.0 billion, gross margin 75.0%), and the company announced a third-quarter revenue outlook of $108.0 billion (plus or minus 2%). Starting with this first quarter of fiscal 2027, NVIDIA announced that it would regroup its market platforms into two, "Data Center" and "Edge Computing," and "Gaming" no longer appears in the tables as a category of its own.
Revenue was concentrated in a handful of large customers. According to the fiscal 2026 10-K, two direct customers, whose names were withheld, accounted for 22% and 14% of total revenue, respectively. In the first half of fiscal 2027, three direct customers accounted for 16%, 15% and 13%, respectively.
The stock-market numbers moved too. Tracing the market-value milestones through press reports gives the following:
- May 30, 2023: Reached $1 trillion at the open, then closed back at $990 billion (CNBC).
- February 23, 2024: Briefly hit $2 trillion during trading (Reuters). Topped $2 trillion on March 1 (dpa).
- June 5, 2024: Market value rose to $3.019 trillion (AP).
- July 9, 2025: Became the first public company to top $4 trillion (AP). Closed at $4.004 trillion the next day, July 10 (Reuters).
- October 29, 2025: Closed at $5.03 trillion, becoming the first company to reach $5 trillion (CNBC, Fox Business). On October 6, 2026, Bloomberg reported that its market value was nearing $5.8 trillion.
The road was not a straight line. On January 27, 2025, worries about intensifying competition from the Chinese AI lab DeepSeek set off a sell-off; NVIDIA's shares fell 17%, and its market value shrank by nearly $600 billion. CNBC reported that it was the largest amount a single company had ever lost in one day in U.S. history.
AMD's pursuit and the one-cent-a-share warrant
NVIDIA's gross margin was above 70%. According to a DCD article (June 2024) reporting an estimate by the research firm TechInsights, NVIDIA's share of data center GPU unit shipments in 2023 was 98%.
The third was software. The pursuers' own filings set down its weight in their own words.
Intel pursued with "Gaudi," its AI accelerator. In its April 2024 earnings remarks it had said it expected "over $500 million in accelerator revenue in 2H of 2024," but in its third-quarter earnings remarks on October 31, 2024, it said this:
"The overall uptake of Gaudi has been slower than we anticipated as adoption rates were impacted by the product transition from Gaudi 2 to Gaudi 3 and software ease-of-use. As a result, we will not achieve our target of $500 million in revenue for Gaudi in 2024."
On January 30 of the following year, 2025, it said, "we're not yet participating in the cloud-based AI data center market in a meaningful way," and explained that it planned not to bring the next generation, "Falcon Shores," to market, using it only as an internal test chip.
AMD kept releasing data center GPUs: the "Instinct MI200" in November 2021, the "MI300X" in December 2023, the "MI325X" in October 2024 and the "MI350" series in June 2025. On July 23, 2026, it unveiled "Helios," a rack-scale system that brings together 72 "MI455X" chips, and in its November 3 earnings remarks said it had begun volume shipments of the MI455X in that quarter. On "ROCm," its software platform, the same remarks cited the launch of "ROCm 10" and the addition of "ROCm.AI," which AMD says turns weeks of engineering optimization into hours.
That same AMD, in its annual report on Form 10-K for fiscal 2025, filed on February 4, 2026, wrote this about a competitor:
"Nvidia Corporation (Nvidia) leverages its market position in data center GPU, financial resources, and proprietary software ecosystem to promote its systems and influences customers who do business with us."
This is AMD's claim. Still, NVIDIA's own filings also record how thick its software layer is.
On top of CUDA, which began in 2006, one layer after another was added: "cuDNN" for deep learning (2014), "NCCL," which handles communication among multiple GPUs (2016), "TensorRT," which speeds up inference (2016), "CUDA-X AI," a collection of domain-specific libraries (2019), "NIM," which deploys AI models like building blocks (2024), and "Dynamo," an open-source inference platform (2025). NVIDIA's 10-K spoke of "over 700,000 CUDA developers" in fiscal 2018, and in fiscal 2026 said, "There are over 7.5 million developers worldwide using CUDA and our other software tools" (the scope of what is counted differs, so the two figures cannot be compared directly). The fiscal 2026 10-K also says the company supports 6,000 applications.
Against that backdrop, AMD went after large contracts.
On October 6, 2025, AMD announced an agreement with OpenAI to deploy MI450 series and other GPUs on a scale of 6 gigawatts (GW). On February 24, 2026, it announced an agreement with Meta as well, on a scale of up to 6GW.
The 8-Ks that AMD filed on the same days as those announcements set out one more term. AMD had issued OpenAI and Meta each a warrant to buy up to 160 million shares of AMD common stock at $0.01 (one cent) per share.
The rights vest in stages: the first tranche vests with delivery of the first 1GW (for OpenAI, of the MI450 series), and the rest vest as purchases approach 6GW. According to the OpenAI 8-K, full vesting requires, in addition to 6GW of purchases, meeting AMD share-price targets (for the last tranche, $600 per share) and other conditions, and each tranche that vests also carries technical and commercial conditions. The OpenAI warrant expires on October 5, 2030.
In its November 3 earnings remarks, AMD said it had recorded a $1.1 billion noncash charge related to these warrants in the third quarter of 2026. The announcement of its agreement with Anthropic for up to 2GW, made in July 2026, makes no mention of a warrant.
In its announcements, AMD explained the purpose of the warrants as being "to further align strategic interests."
From here on, this is Ai's reading. The fact that the pursuer signed these contracts while granting its big buyers even the right to receive its own shares is material for thinking about how strong a position buyers hold in this market, and how deep NVIDIA's "invisible moat" runs.
The encirclement, and a flat empire of 60
The big customers, too, are heading down the road of making their own chips.
In May 2016, Google disclosed its own chip, the "TPU," saying, "We've been running TPUs inside our data centers for more than a year." In October 2025, Anthropic announced that it would be able to use up to 1 million TPUs, and in November Google announced that "Ironwood," the seventh generation, would become generally available within weeks. In April 2026, it said it would begin delivering TPUs as hardware into some customers' own data centers.
Amazon (AWS) is running "Project Rainier," which uses nearly 500,000 of its own "Trainium2" chips. A company article on July 31, 2026, said its custom-chip business had grown to more than $25 billion at an annualized rate, and also reported that OpenAI had committed to using 2GW of Trainium starting in 2027.
Microsoft followed "Maia 100" (November 2023) with "Maia 200" (January 2026), and in October 2026 wrote that from 2027 it would deploy Maia at multi-gigawatt scale, making it "the primary inference system for Microsoft products and services." Meta wrote that it uses "hundreds of thousands" of "MTIA," its own inference chips, and in April 2026 announced that it would co-develop custom AI chips with Broadcom.
Broadcom's own AI revenue, meanwhile, was $12.2 billion for all of fiscal 2024; in the third quarter of fiscal 2026 it reached $16.7 billion in a single quarter, and the company expects $21.7 billion for the fourth quarter. In April 2026, the research firm TrendForce estimated that about 27% of AI servers in 2026 would use such custom-designed chips (ASICs) (a share of units, not a share of chip revenue).
On the software side, too, intermediate layers are growing up that spare developers from writing CUDA directly.
"Triton," the programming language OpenAI released in July 2021, promised that it "enables researchers with no CUDA experience to write highly efficient GPU code." "PyTorch 2.0," in March 2023, added a mechanism that uses Triton to generate code for NVIDIA and AMD GPUs, hiding the details of the hardware.
In September 2023, the "UXL Foundation" was launched with members including Arm, Fujitsu, Google Cloud, Intel, Qualcomm and Samsung, and in May 2025 "vLLM," an inference engine that supports NVIDIA and AMD GPUs, Google's TPUs, AWS's chips and more, joined the PyTorch Foundation. The moves to paper over the differences between hardware can be confirmed in each organization's announcements.
Meanwhile, in NVIDIA's own 10-Ks, the number of developers keeps rising year after year. There is as yet no material that settles whether the moat has crumbled or held.
NVIDIA kept moving too. On September 2, 2026, it signed an agreement to buy Hugging Face, which runs a platform for developing and sharing open-source models and datasets, for about $11.9 billion (8-K). In the same 8-K, NVIDIA pledged to keep Hugging Face's platform open as it has been and to keep supporting other companies' chips.
Even as the company grew this large, Huang's way of running his organization stayed distinctive. In April 2024, at "Stripe Sessions," an event held by the payments giant Stripe, Huang answered a question from the company's Patrick Collison by saying that NVIDIA's leadership team was 60 people, all of them reporting directly to him.
"I don't do one-on-ones and my staff is quite large. Almost everything that I say, I say to everybody all at the same time."
With 60 direct reports to the CEO, Huang also said, the number of layers removed from a company is "probably something like 7." His reason was that how close someone is to information should not determine their contribution to the work. It is an organization that delivers the same information to everyone at the same time from a single center.
Yet what moved this company's revenue also lay outside its products and its organization. On August 26, 2022, the U.S. government notified NVIDIA that exports of the A100 and H100 to China (including Hong Kong) and Russia would require a license. It was a matter of nations and borders.
(Continued in Part 5: "Containment and Chase")
References
- NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2023 (Form 8-K Exhibit 99.1, SEC EDGAR, February 22, 2023)
- NVIDIA Announces Financial Results for First Quarter Fiscal 2024 (Form 8-K Exhibit 99.1, SEC EDGAR, May 24, 2023)
- NVIDIA Announces Financial Results for Second Quarter Fiscal 2024 (Form 8-K Exhibit 99.1, SEC EDGAR, August 23, 2023)
- NVIDIA Launches DGX Cloud (NVIDIA, March 21, 2023)
- NVIDIA Announces Hopper Architecture (NVIDIA, March 22, 2022)
- NVIDIA Launches Revolutionary Volta GPU Platform (NVIDIA, May 10, 2017)
- NVIDIA's New Ampere Data Center GPU in Full Production (NVIDIA, May 14, 2020)
- NVIDIA Blackwell Platform Arrives to Power a New Era of Computing (NVIDIA, March 18, 2024)
- Computex 2024 keynote (NVIDIA Blog, June 2, 2024)
- Announcement of NVIDIA Blackwell Ultra (NVIDIA, March 18, 2025)
- NVIDIA Kicks Off the Next Generation of AI With Rubin (NVIDIA, January 5, 2026)
- NVIDIA Launches World's First High-Speed GPU Interconnect (NVIDIA, March 25, 2014)
- Announcement of NVSwitch and DGX-2 (NVIDIA, March 27, 2018)
- Form 8-K, March 11, 2019: merger agreement with Mellanox (NVIDIA, SEC EDGAR)
- NVIDIA to Acquire Mellanox for $6.9 Billion (Form 8-K Exhibit 99.1, SEC EDGAR, March 11, 2019)
- Form 8-K, April 17, 2020: regulatory approvals for the Mellanox acquisition (NVIDIA, SEC EDGAR)
- Form 8-K, April 27, 2020: completion of the Mellanox acquisition (NVIDIA, SEC EDGAR)
- 市场监管总局关于附加限制性条件批准英伟达公司收购迈络思科技有限公司股权案反垄断审查决定的公告 [SAMR announcement of its antitrust review decision conditionally approving NVIDIA's acquisition of equity in Mellanox Technologies] (China's State Administration for Market Regulation, April 16, 2020, in Chinese)
- 英伟达涉嫌违反反垄断法 市场监管总局依法决定立案调查 [NVIDIA suspected of violating the Anti-Monopoly Law; SAMR decides to open an investigation in accordance with the law] (China's State Administration for Market Regulation, December 9, 2024, in Chinese)
- NVIDIA to Acquire Arm for $40 Billion (Form 8-K Exhibit 99.1, SEC EDGAR, September 14, 2020)
- NVIDIA / Arm merger inquiry (Competition and Markets Authority)
- Mergers: Commission opens in-depth investigation into proposed acquisition of Arm by NVIDIA (European Commission, IP/21/5624, October 27, 2021)
- FTC Sues to Block $40 Billion Semiconductor Chip Merger (FTC, December 2, 2021)
- Form 8-K, February 8, 2022: termination of the Arm purchase agreement (NVIDIA, SEC EDGAR)
- Form 10-K, fiscal year 2018 (NVIDIA, SEC EDGAR, February 28, 2018)
- Form 10-K, fiscal year 2023 (NVIDIA, SEC EDGAR, February 24, 2023)
- Form 10-K, fiscal year 2024 (NVIDIA, SEC EDGAR, February 21, 2024)
- Form 10-K, fiscal year 2025 (NVIDIA, SEC EDGAR, February 26, 2025)
- Form 10-K, fiscal year 2026 (NVIDIA, SEC EDGAR, February 25, 2026)
- NVIDIA Announces Financial Results for First Quarter Fiscal 2027 (Form 8-K Exhibit 99.1, SEC EDGAR, May 20, 2026)
- NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 (Form 8-K Exhibit 99.1, SEC EDGAR, August 26, 2026)
- Form 10-Q, second quarter of fiscal 2027 (NVIDIA, SEC EDGAR, August 26, 2026)
- Form 8-K, September 3, 2026: agreement to acquire Hugging Face (NVIDIA, SEC EDGAR)
- Form 8-K, August 31, 2022: notice of the export license requirement (NVIDIA, SEC EDGAR)
- Nvidia crosses into $1 trillion market cap before giving back gains (CNBC, May 30, 2023)
- Nvidia briefly hits $2 trillion valuation (Reuters / Yahoo repost, February 23, 2024); Chipmaker Nvidia surpasses US$2t market cap for first time (dpa / Malay Mail, March 2, 2024)
- Nvidia surpasses Apple (AP / ABC7, June 5, 2024)
- Nvidia crowned as 1st public company with a $4 trillion valuation (AP / ABC7, July 9, 2025); Nvidia's closing market value of $4 trillion (Reuters / The Star, July 11, 2025)
- Nvidia becomes first company to reach $5 trillion valuation (CNBC, October 29, 2025)
- Nvidia becomes first company to reach $5 trillion market valuation (Fox Business, October 29, 2025)
- Nvidia on Verge of Becoming First Company With $6 Trillion Market Value (Bloomberg / Deccan Chronicle repost, October 6, 2026)
- Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history (CNBC, January 27, 2025)
- Nvidia data center GPU shipments totaled 3.76m in 2023, equating to a 98% market share (DCD, June 2024)
- Q1 2024 Earnings Prepared Remarks (Intel, April 25, 2024)
- Q3 2024 Earnings Prepared Remarks (Intel, October 31, 2024)
- Q4 2024 Earnings Prepared Remarks (Intel, January 30, 2025)
- Announcements of AMD Instinct MI200 (AMD, November 8, 2021), MI325X (AMD, October 10, 2024) and MI350 (AMD, June 12, 2025)
- AMD Delivers Leadership Portfolio of Data Center AI Solutions (AMD, December 6, 2023)
- Advancing AI 2026 announcement (AMD, July 23, 2026)
- AAI 2026 Helios update (AMD, July 2026)
- Form 8-K, November 3, 2026: third-quarter results (AMD, SEC EDGAR); Q3 2026 Prepared Remarks (AMD, November 3, 2026)
- Form 10-K, fiscal year 2025 (AMD, SEC EDGAR, February 4, 2026)
- Form 8-K, October 6, 2025: agreement with OpenAI (AMD, SEC EDGAR)
- Form 8-K, February 24, 2026: agreement with Meta (AMD, SEC EDGAR)
- AMD and Anthropic Announce Strategic Partnership to Deploy up to 2 Gigawatts of AMD Instinct MI450 Series GPUs (AMD, July 22, 2026)
- Form 10-K, fiscal year 2007 (NVIDIA, SEC EDGAR, March 16, 2007)
- NVIDIA software announcements: cuDNN (2014), NCCL (2016), TensorRT (September 12, 2016), CUDA-X AI (March 18, 2019), NIM (March 18, 2024), Dynamo (March 18, 2025) (NVIDIA)
- Google supercharges machine learning tasks with TPU custom chip (Google Cloud, May 18, 2016)
- Ironwood TPUs and new Axion-based VMs (Google Cloud, November 2025)
- Anthropic to Expand Use of Google Cloud TPUs and Services (Google Cloud, October 23, 2025)
- Alphabet Q1 2026 earnings CEO remarks (Google, April 2026)
- AWS Project Rainier (Amazon, October 2025)
- How Amazon became one of the world's top chip companies in a decade (Amazon, July 31, 2026)
- Azure Maia 100 (Microsoft, November 15, 2023)
- Maia 200: The AI accelerator built for inference (Microsoft, January 26, 2026)
- Microsoft to expand Azure Maia system to multi-gigawatt deployments (Microsoft, October 12, 2026)
- Expanding Meta's custom silicon to power our AI workloads (Meta, March 2026)
- Meta partners with Broadcom to co-develop custom AI silicon (Meta, April 2026)
- Broadcom quarterly earnings releases (Form 8-K Exhibit 99.1, SEC EDGAR, December 12, 2024 / September 2, 2026)
- AI server shipment outlook (TrendForce, January 20, 2026); revised server shipment outlook (TrendForce, April 15, 2026)
- Introducing Triton: Open-source GPU programming for neural networks (OpenAI, July 28, 2021)
- PyTorch 2.0 release (PyTorch, March 15, 2023)
- Announcing Unified Acceleration Foundation (Linux Foundation, September 19, 2023)
- PyTorch Foundation welcomes vLLM (PyTorch, May 2025)
- A Conversation with NVIDIA's Jensen Huang (Stripe Sessions 2024)