European and American Chip Giants: The "Shovel Seller" Chess Game in the AI Era
The most crowded track in the global capital market is not any single AI application, but the group of companies that build AI — chips. While OpenAI, Anthropic, xAI and others are competing for a "trillion-dollar valuation" at the front stage, the few European and American chip giants behind them that control GPUs, fabs and lithography machines are the ones making steady profits. They do not develop large models, but define how fast, how expensive, and how safe large models can run. This is a power pyramid belonging to the "shovel sellers".
While OpenAI, Anthropic, and xAI are competing for trillion-dollar valuations at the front stage, the European and American chip companies along the industrial chain that provide GPUs, wafer foundry services and lithography equipment are the ones making truly stable profits. They do not develop large models, but determine how fast, how expensive, and how safe large models can run.
"Shovel seller" is an old metaphor. During the California Gold Rush in 1849, the people who made steady profits were often those who sold shovels, jeans and tickets. The 2026 AI Gold Rush has pushed this logic to the extreme: Microsoft, Google, Amazon, Meta, and Oracle are investing hundreds of billions of dollars each year to build data centers, and the largest share of their expenditure flows to European and American semiconductor companies. Gold miners face fierce competition, while shovel sellers almost take all the profits.
If we compress the 2026 global semiconductor industry into a single diagram, we can see a pyramid-style power chain: at the top is design IP and architecture (Arm), followed by GPU/ASIC design (NVIDIA, AMD, Broadcom), then wafer manufacturing (TSMC, Intel), and the base that supports the entire pyramid is the lithography machine that only one company can build (ASML). Along this chain, European and American companies occupy every irreplaceable link from architecture to equipment. Outside this chain, there is a unique "outlier" — Apple does not sell any chips externally, but consumes a large amount of the world's advanced process capacity every year, participating in the computing power race in its own way.
NVIDIA: The Super Winner at the Top of the Pyramid
NVIDIA remains the company with the largest size and strongest bargaining power on this chain. As of mid-October, its market value is about 5.7 trillion US dollars, making it the most valuable company in the world, with a size equivalent to the sum of several tech giants behind it. Its revenue in the latest fiscal quarter (as of July 26, 2026) reached 96.2 billion US dollars, a year-on-year increase of 106%, of which the data center contributed 89 billion US dollars, accounting for about 92.5%; the gross profit margin was about 75%. The company expects revenue in the next fiscal quarter to be about 108 billion US dollars, with a year-on-year increase of about 89%. In the last fiscal year (as of January 2026), its revenue was 215.9 billion US dollars, the data center contributed 193.7 billion US dollars, and the net profit was 120.1 billion US dollars.
NVIDIA's core barrier is not just chips, but CUDA. Over the past two decades, a large number of scientific research papers, frameworks, and operators run on CUDA by default. Switching platforms means rewriting code, retraining models, and re-adjusting performance. AMD's hardware parameters can get close, but it has always been difficult to shake NVIDIA's market share — what customers buy is not just a graphics card, but the overall availability of "graphics card + software + ten years of accumulation".
In terms of product rhythm, Blackwell (B200/GB200) has been fully ramped up, and the next-generation Vera Rubin NVL72 has also entered the accelerated mass production stage in mid-2026. The company expects that orders for the Blackwell and Rubin platforms will accumulate to at least 1 trillion US dollars by 2027. At the end of September, NVIDIA approved an additional 1500 billion US dollar stock repurchase plan, and the total remaining repurchase authorization reached 2350 billion US dollars.
The risks are also clear: the export control of H20 to the Chinese market caused a 4.5 billion US dollar impairment in the last fiscal quarter; the longer-term threat comes from customers' self-developed ASICs — Microsoft, Google, Amazon, and Meta not only buy its chips, but are also preparing to "de-NVIDIAize". But at least for now, no competitor can approach it in terms of both "scale + ecosystem".
TSMC: The Underestimated "Geopolitical Pricing Power"
TSMC (Taiwan, China) is the most special existence on this chain — it does not design chips, but manufactures the vast majority of the world's advanced process chips. In the second quarter of 2026, its revenue was 40.2 billion US dollars, a year-on-year increase of 33.7%, the net profit margin was 55.6%, and the gross profit margin was 67.7%; the revenue from 7nm and more advanced processes accounted for 77% of wafer revenue, of which 2nm has begun to contribute revenue (3%), 3nm accounts for 30%, and 5nm accounts for 33%. The revenue in the third quarter is expected to reach a maximum of 45.8 billion US dollars. CFO Huang Renzhao clearly stated that 2nm is becoming a new revenue growth engine, and production will be accelerated in the third quarter.
Its moat is about 70% of the advanced process foundry share, and more than 90% of the world's leading node production capacity. More critically, it has "geopolitical pricing power": under uncertainty, TSMC has instead become a "silicon shield" that no party dares to lose. The total investment in the Arizona plant has risen to 265 billion US dollars, with a planned cluster of six factories. The first plant has achieved profitability, which is narrowing the "Taiwan discount" that has long weighed on its stock price. Under the leadership of C.C. Wei, TSMC translates AI demand into solid cash flow with one fab after another.
But TSMC's risk also lies here: the more concentrated and critical the production capacity is, the more it becomes a "single point of failure" in geopolitics. Once the situation across the Taiwan Strait changes, 90% of the global advanced chip supply will evaporate in an instant. This is the underlying logic behind the US's push for "capacity diversification", and it is also the contradiction where TSMC is forced to expand production overseas and dilute its own moat.
ASML: Europe's Only EUV Supplier
If the United States dominates the field of chip design, the hardest bargaining chip for Europe in the semiconductor chain is ASML from the Netherlands — the world's only EUV lithography machine supplier. Without its machines, there would be no advanced chips below 7nm. This "one and only" position makes ASML the scarcest fulcrum for Europe on the global tech power map.
ASML's current market value is about 672 billion US dollars, with an increase of about 50% during the year. The market is optimistic about its third-quarter performance, expecting sales to reach 11.5 billion euros, supported mainly by AI-driven demand for advanced chips. The more critical variable is High-NA EUV: the company is cooperating with Zeiss to promote the R&D of Hyper-NA EUV. Goldman Sachs estimated in a recent research report that the annual revenue of High-NA is expected to exceed 15 billion euros by 2028 — this is the ticket to the sub-2nm and even 1nm era.
In terms of risks, ASML's customers are highly concentrated, with customers in Taiwan region contributing about a quarter of its revenue, and TSMC is its largest single customer. This "interdependence" itself is a geopolitical vulnerability: when TSMC is booming, ASML reaps the benefits; when the Taiwan Strait situation changes, ASML's orders and geopolitical narratives will be under pressure at the same time. Monopoly brings huge profits, but also brings structural risks of "success depends on customers, crisis also depends on customers".
Broadcom: The Hidden Winner of Custom Chips
Broadcom is an underestimated AI chip giant. In early September 2026, its market value was about 1.75 trillion US dollars, making it the second most valuable semiconductor company after NVIDIA, but the least talked about by ordinary investors. In its latest fiscal quarter (as of August 2), revenue was 29.6 billion US dollars, a year-on-year increase of 86%, and AI semiconductor revenue reached 16.7 billion US dollars, a year-on-year increase of 221%, accounting for about 56% of total revenue. The company expects revenue in the fourth fiscal quarter to be about 34.8 billion US dollars, and AI semiconductor revenue will further rise to 21.7 billion US dollars, a year-on-year increase of 236%.
Hock Tan's strategy is not to sell general-purpose GPUs, but to customize XPU/ASIC for six major customers including Google, Meta, OpenAI, and Anthropic, and sell network chips that connect thousands of chips into clusters. The shrewdness of this path is: it does not bet on the victory or defeat of a certain general architecture, but turns "I will build whatever customers want" into a business — the more chips customers develop on their own, the more "design + network" fees Broadcom collects. The management set a target of AI chip revenue exceeding 100 billion US dollars in FY2027. Broadcom does not appear in the "living room", but hides deep in every data center that trains large models.
AMD: The Alternative Option Beyond NVIDIA
AMD is the player on this chain that most resembles a "head-on challenger to NVIDIA". On September 21, 2026, AMD's market value exceeded 1 trillion US dollars for the first time, and its stock price rose by 185% during the year. Data center revenue in the second quarter reached 6.7 billion US dollars, a year-on-year increase of 107%, and data center operating profit turned to a profit of 2.1 billion US dollars from a loss of 155 million US dollars in the same period last year.
The MI350 accelerator series is being used by hyperscale customers as a diversified choice beyond NVIDIA. Lisa Su is slowly but firmly eroding CUDA's market share with the ROCm software stack. AMD also disclosed multi-year agreements with OpenAI and Meta, with each party planning to deploy up to 6 GW of AMD data center GPUs.
AMD's situation is a typical "second growth curve" — it replicates the "cost-effectiveness + open ecosystem" strategy that once defeated Intel in the CPU battlefield to the GPU/accelerator field. Although the revenue of data center GPUs is only a fraction of NVIDIA's, global cloud vendors and large enterprises will inevitably give AMD at least one verification opportunity for the risk control purpose of "de-single supplier".
Its ceiling is also clear: the catch-up of the software ecosystem is not only a matter of ROCm maturity, but also that CUDA has been deeply embedded in scientific research and teaching, forming developer inertia. Therefore, the sign of AMD's success is not defeating NVIDIA, but firmly occupying the position of "the second computing power pillar" in customers' data centers.
Apple: The Closed-loop Player with Self-developed and Self-used Chips
Apple does not sell any chips externally, but consumes a large amount of the world's advanced process capacity every year. It is one of the first customers of TSMC's 3nm/2nm nodes and occupies an important share in the initial 2nm production capacity. Apple's self-developed chips cover all core computing units of iPhone, Mac, iPad and Vision Pro, and its market value remains at around 4.5 trillion US dollars.
Apple's killer move is the computing power overflow of vertical integration. In early 2026, private cloud computing servers were directly upgraded from M2 Ultra to M5 chips, skipping the M3 and M4 generations; the M5 Ultra server with code name "J246" is in early development, and the AI inference server for enterprise customers is planned to be equipped with two or four self-developed Ultra chips, which will be launched as early as 2029. Apple is also discussing with NVIDIA the use of NVLink Fusion for server chip interconnection. The annual shipment of 200 million mobile phones and 20 million Macs dilutes R&D costs, and the software ecosystem (Metal, Core ML) is deeply bound to hardware.
But the closed loop is also its ceiling: it cannot enjoy the premium from external customers like NVIDIA, and has no say in the cloud AI training side. In 2026, its relationship with OpenAI changed and it reached an agreement with Google Gemini, exposing its shortcomings in cloud AI capabilities. Apple is betting on "device-cloud collaboration" — devices are responsible for perception and light inference, and the cloud is responsible for heavy models, and it wants to occupy positions on both ends.
Intel: The Process Chaser Betting on 18A and 14A
Intel's story is the most ups and downs. It once lost 18.8 billion US dollars in 2024, and its net loss narrowed significantly in 2025. Its stock price rebounded sharply from the low point in 2026, its market value once exceeded 560 billion US dollars, and it is currently about 470 billion US dollars. New CEO Lip-Bu Tan defined 2026 as the year of execution and breakthrough: the 18A process (about 2nm level) has been mass-produced since October 2025, and Panther Lake was launched in March 2026. According to the company's disclosure, the yield in the second quarter has reached about 80%.
Positive signals have emerged in the foundry business: foundry revenue in the second quarter was 5.8 billion US dollars, driven by strong growth in 18A output — according to the company's disclosure, 18A output exceeded the target by about 25%, with a sequential increase of more than 50%. External foundry revenue was 293 million US dollars, and the operating loss of the foundry business improved by 730 million US dollars quarter-on-quarter. Intel has raised its 2026 capital expenditure to more than 20 billion US dollars, and expects 2027's capital expenditure to be significantly higher than 2026, with the vast majority invested in the United States. 14A is planned to be mass-produced in 2029, and will adopt High-NA EUV for the first time.
Intel is betting on the dual dividends of politics and industry brought by "U.S. local manufacturing + process overtaking" — the CHIPS Act subsidies and customers' demand for "non-TSMC" production capacity are all on its side. If it succeeds, it will return to the first echelon; if it fails, it will continue to be suppressed by TSMC.
Computing Power Investment: Capital Expenditure and Chip Orders Reinforce Each Other
To understand why these seven companies can earn exaggerated profits, we cannot avoid an underlying variable: global AI computing power investment is expanding at the pace of an "investment race". The combined capital expenditure of Microsoft, Google, Amazon, and Meta, the four hyperscalers, has risen to 725 billion US dollars in 2026, higher than the previous market expectation of 670 billion US dollars; among them, Microsoft's expected capital expenditure is about 190 billion US dollars, and Meta has raised its capital expenditure to 125-145 billion US dollars. This forms a cycle: the more advanced the chip, the larger the model can be; the larger the model, the bolder customers are to place larger orders; larger orders in turn justify the expansion of chip fabs.
But the other side of this card is the risk of "supply-demand mismatch". Today's order boom is built on the assumption that "the return on AI investment will eventually be realized". Once a hyperscaler finds that the utilization rate of self-built computing power is lower than expected, or cuts external procurement after self-developed ASICs mature, these "shovel sellers" with extremely high expectations will be the first to bear the brunt. The 2026 chip prosperity is not only a technological victory, but also a fragile balance built on high concentration and high assumptions.
Arm and Qualcomm: The Tax Collectors of Architecture and Terminals
Arm, based in Cambridge, UK, is the light-asset winner on this chain — it does not make chips, only licenses CPU architectures and collects royalties. Its FY2026 revenue was 4.92 billion US dollars, a year-on-year increase of 23%, with a gross profit margin as high as 97% to 98%. With the popularization of Armv9 and Neoverse's penetration rate in the hyperscale data center CPU market approaching 50%, Arm is transitioning from "mobile phone architecture" to "the logical language of the AI era". Its first self-developed AGI CPU has been adopted by customers such as Meta and Oracle, and more than 2 billion US dollars of customer orders have been recorded in FY2027 and FY2028, but its current market value fluctuates sharply.
Qualcomm focuses on end-side AI and automotive chips, etc. Its FY2025 revenue was about 44.3 billion US dollars, and it is a key part of "AI landing on mobile phones and cars". According