In the eye of the global AI computing power competition, Marvell, Broadcom, and Google are engaged in a covert battle using ASIC chips as their weapon, determining the technological hegemony for the next decade. This war, devoid of gunpowder, not only affects the strategic nerves of tech giants like NVIDIA, Microsoft, and Meta but also approaches a pivotal turning point in 2025—the fate of supercomputing chips is quietly unfolding on the North American continent.
Marvell’s Life-and-Death Gamble: A Technological Odyssey of Three Generations
As the core supplier of AWS supercomputing chips, Marvell is facing the most complex technological iteration in its history: its Tranium2 chip is entering the end of production in 2025, while the transitional product Tranium2.5 is caught in a tug-of-war over packaging solutions. Data shows that Marvell can only handle about 200,000 units per quarter of Tranium2, which is less than one-third of AWS’s total demand. To validate the feasibility of the new packaging, AWS has even allocated part of Tranium2.5’s production capacity to Marvell as a testing ground. This strategy of “building guns while fighting” reflects AWS’s concerns about Marvell’s technological capabilities.
The real highlight will be in 2026: the Tranium3, utilizing TSMC’s CoWoS-L packaging, is set to enter mass production, with design complexity that is among the highest in the industry—combining 4 computing die and 4 I/O die requires over 12 reroute layers and more than 2000 microbump connections. A more severe challenge comes from the supply chain: if the joint design between Marvell and Annapurna encounters yield issues in Q2 2025, AWS may transfer up to 500,000 units of Tranium3 capacity to Broadcom, which would directly result in Marvell losing its largest customer, AWS.
Meanwhile, the Tranium4, codenamed “Maverick,” has quietly begun its design phase. This chip, utilizing TSMC’s N3P process, is planned for mass production in Q4 2027, with performance targets aimed at three times that of the current flagship product. However, whether Marvell can reach the finish line depends on the yield of Tranium2.5 in Q4 2025—this race against time and technology has entered its final sprint.
Broadcom’s Opportunistic Entry: Ambitions for Computing Power Amidst a Merger Frenzy
As Marvell grapples with technological bottlenecks, Broadcom is aggressively expanding its ASIC territory through acquisitions. From acquiring Marvell’s networking chip business to swallowing SiFive, this chip giant has built a complete product line covering cloud AI training, inference, and edge computing. Its latest “Titan” series AI chips, utilizing TSMC’s CoWoS-S packaging, are planned for mass production in 2026, targeting the training chip market dominated by OpenAI.
What is even more concerning is Broadcom’s deep integration with Google: in the Google TPU V7 project, Broadcom is not only responsible for design but also undertakes 80% of the manufacturing tasks. According to supply chain sources, Google plans to procure 4 million TPU V7 units in 2026, with 3 million units for its own AI training and the remaining 1 million for external customers. This “self-research + foundry” model significantly enhances Broadcom’s bargaining power in the AI chip market.
In the Apple camp, Broadcom’s ASIC solutions for Siri voice chips have achieved breakthroughs in energy efficiency at the 12nm process, with AI inference speeds improved by 300% compared to previous products. This technological advantage is prompting Apple to consider shifting more core chips to Broadcom’s designs, which would undoubtedly be a fatal blow to Marvell.
Google’s Computing Power Empire: Vertical Monopoly from TPU to Optical Modules
As the only player among the three North American supercomputing giants to self-develop and self-use, Google’s ASIC strategy exhibits remarkable coherence: from the first-generation Axion CPU to the second-generation Tamar CPU, and now to the latest TPU codenamed “V7,” its product line covers all scenarios from general computing to AI training. In 2026, Google plans to increase TPU production capacity to 4 million units, with 2 million units for internal AI training, 1 million units for sale to AI startups like Anthropic, and the remaining 1 million units for building the world’s largest AI computing cloud.
But Google’s ambitions do not stop there. In the optical module sector, its custom 1.6T optical module demand is expected to surge to 6 million units in 2026, accounting for nearly 60% of global capacity. This absolute control over the upstream supply chain gives Google a cost advantage in the AI computing power competition—internal data shows that the unit computing cost of its self-developed TPU is 40% lower than that of NVIDIA’s GPUs.
More aggressive strategies are reflected in talent acquisition: Google has poached over 200 senior chip designers from Marvell and Broadcom, forming an ASIC R&D team of 1,500 people. This “technology + talent + supply chain” triad strategy positions Google to potentially become the world’s largest AI chip manufacturer by 2027, with a market share exceeding 25%.
Life-and-Death Speed: The Critical Three Years from 2025 to 2027
The outcome of this supercomputing chip war will be revealed in the next three years:
– Q4 2025: The yield of Marvell’s Tranium2.5 will determine whether it can retain AWS orders; if the yield is below 70%, Broadcom will take over the majority of Tranium3’s capacity;
– Q2 2026: The yield ramp-up speed of Google TPU V7 will determine whether it can achieve its shipment target of 4 million units, which directly relates to whether its AI training cost can be reduced to $0.5 per PFlop/s;
– Q1 2027: The production progress of Meta’s Arke chip will determine whether it can capture a 15% share of the inference chip market; otherwise, its AI training costs will be 35% higher than Google’s.
In this trillion-dollar computing power competition, there are no permanent winners. Can Marvell break through the packaging technology bottleneck? Will Broadcom’s mergers and integrations yield synergies? Will Google’s vertical monopoly trigger antitrust investigations? The answers to these questions will determine who can ascend to the AI computing throne in 2030.
Scan the code to join us and learn more about investment opportunities in advance.
