In a recent earnings conference call, executives from tech giant Amazon expressed ambitious plans to invest up to $100 billion this year. The bulk of this investment is aimed at building data centers, collaborating with chip manufacturers to create AI chips, and making substantial equipment investments. This is all in the pursuit of providing AI computing resources. However, they cautioned that their cloud computing division, Amazon Web Services (AWS), might still face capacity constraints, which could hinder their ability to meet the surging demand for AI computing power from cloud customers. This caution comes amid a trend where major tech players in the U.S. are pouring money into the AI sector, indicating that two major AI ASIC (Application-Specific Integrated Circuit) companies might emerge as significant beneficiaries in this cash-burning frenzy.

CEO Andy Jassy’s vision for Amazon includes transforming it into an "AI superstore" by making significant investments to maintain the company’s leading position in cloud computing services. However, Jassy warned of the possibility of “significant volatility” in growth, suggesting that delays in hardware procurement and inadequate power supply could pose challenges in scaling their AI capabilities.

During the conference following the release of their Q4 financial report, Jassy stated, “If there were no capacity constraints or limits in production, our growth could be much faster than anticipated.” His concerns about AI capacity echo similar sentiments from industry giants Microsoft, Meta, and Google, all of whom acknowledge that they are struggling to satisfy surging AI computing needs. AWS and Microsoft’s Azure dominate the cloud computing market, together accounting for over 50% of the sector.

Last week, Microsoft pointed out that their cloud business revenue growth has been significantly negatively impacted due to a lack of sufficient data centers to meet the enormous AI computing needs of their developers and cloud inference endpoints.

As the landscape changes with innovations such as DeepSeek’s low-cost computing paradigm, AI training and application inference costs are set to decline markedly. Under a modest $6 million investment, DeepSeek has demonstrated the ability to develop AI models that compete with OpenAI’s offerings using chips that greatly underperform compared to current industry standards.

Despite these advancements, the latest financial results demonstrate that tech giants like Amazon, Microsoft, and Google remain committed to hefty spending in AI. The rationale is that they are betting on the new low-cost computing paradigm propelling AI applications across diverse industries, resulting in an exponential increase in demand for inference-level AI computing resources. This explains why companies like ASML emphasize that declining AI costs will likely expand the scope of AI applications significantly.

Looking at the recent shifts in global investment and stock market dynamics, it appears that the biggest beneficiaries from the massive AI expenditures of American tech firms are not the traditional “AI chip overlord” Nvidia, but rather AI ASIC leaders Broadcom and Marvell.

The central logic behind this latest investment trend is that as generative AI software and AI agents become more widely adopted, the demand for AI computing power at the cloud inference level will dramatically increase. Combined with DeepSeek's groundbreaking approach that significantly reduces inference costs, the self-developed AI ASICs created in partnership with Broadcom or Marvell are poised to outperform Nvidia’s AI GPUs in terms of performance, cost, and energy consumption in mass-scale neural network parallel computing.

As a result, the rapid growth of the inference AI chip market presents substantial opportunities for companies like Broadcom and Marvell to expand their market share in the evolving landscape, potentially mirroring the explosive growth trajectory often associated with Nvidia.

To realize their vision of a robust AI future, American tech giants are committed to continued substantial financial investments.

Jassy indicated that supply issues related to AI chips—whether sourced from third parties like Nvidia or developed internally at Amazon—along with power supply bottlenecks, are constraining AWS’s ability to bring newly established large data centers online. He mentioned that these constraints might ease in the latter half of 2025 as resources become integrated into AI projects.

In the last quarter of 2024, Amazon's capital expenditure reached approximately $26.3 billion, with the majority focused on AI-related projects within the AWS division and favoring self-developed ASICs over purchasing Nvidia’s AI GPUs. Jassy assured analysts that this spending figure is a reasonable representation of the company's anticipated expenditure pace for 2025.

According to their quarterly financial report, AWS reported a remarkable revenue increase of 19% in the quarter ending December 31, reaching $28.8 billion. This marks the third consecutive quarter of growth at or above 19% for the cloud division. Notably, AWS's operating profit soared to $10.6 billion, surpassing broader market expectations of $10 billion and achieving a year-on-year surge of 47%, reflecting an expanding customer base and escalating demand for AWS's AI application software development platform—Amazon Bedrock—which aims to streamline the deployment of AI applications and provide essential AI inference computing resources.

As AWS continues its upward trajectory, it remains a crucial profit center within Amazon.

Despite the growth, eMarketer analyst Sky Canavis pointed out that AWS's expansion has stabilized rather than accelerated, indicating that Amazon faces similar AI computing resource constraints as its competitors Google and Microsoft.

As of the close of trading in New York on Friday, Amazon shares were priced at $238.83. Following the earnings outlook that fell short of expectations, Amazon saw a drop of over 4% in after-hours trading. Thus far in the year, Amazon's stock has increased by 8.9%, with a substantial rise of 44% projected for 2024.

Analysts have voiced concerns about the potential impact of the “AI cash burn competition” on profitability. In their outlook, Amazon forecasted operating profits between $14 billion and $18 billion for the quarter ending in March, which is lower than analysts’ average expectations of $18.2 billion; they anticipate total revenues for the quarter could reach as high as $155.5 billion, falling short of the average projection of around $158.6 billion.

“Although Amazon's overall quarterly performance is encouraging, investor focus has promptly shifted to the first quarter's guidance, which didn't meet expectations due to exchange rate impacts and spending constraints,” shared DA Davidson & Co. analyst Gill Luria.

Amid the substantial decline in AI training costs led by DeepSeek, coupled with reduced costs for inference tokens, AI agents and generative AI software are poised to infiltrate various sectors rapidly. Responses from Western tech giants such as Microsoft, Meta, and ASML have shown appreciation for DeepSeek’s innovations, yet their determination to continue large-scale AI investments remains unshaken. These corporations believe that the technology path introduced by DeepSeek promises a general decline in AI costs, leading to greater opportunities and significantly more substantial demand for AI applications and inference-level computing resources.

Regarding the spending plan for 2025, Amazon’s management anticipates it will reach up to $100 billion, emphasizing their belief that the emergence of DeepSeek heralds an era of significant expansion in inference-level AI computing demand. Consequently, increased investments are being funneled into supporting AI business growth. Jassy noted, “We won’t engage in procurement without seeing significant demand signals. When AWS raises its capital expenditure, especially in opportunities as rare as AI, I perceive this as a positive sign for the long-term development of AWS.”

Last week, Google, Microsoft, and Meta reiterated their commitment to funneling substantial resources into the AI space. Despite encountering the low-cost disruption posed by DeepSeek, these tech giants remain confident that these large-scale investments will lay a solid foundation for future AI computing demands.

According to Visible Alpha forecasts, Microsoft’s capital expenditure for 2025 is projected to exceed $90 billion, representing over 30% of their revenue. Facebook parent company Meta has also significantly ramped up its investment plans, recently announcing a more than 60% increase in capital expenditures for 2025, reaching up to $65 billion, also exceeding 30% of their revenue. This follows Meta's extravagant spending of over $38 billion in 2024 on cutting-edge technologies such as AI. Furthermore, Google plans to invest $75 billion in 2025 into projects related to AI, marking a substantial rise from last year's $52.5 billion investment, significantly higher than the average under 13% over the past decade.

As discussions unfold around the “cash-burning craze” led by tech giants, the market is positioning AI ASIC manufacturers—like Broadcom and Marvell—as potential major winners in this race.

With American tech giants pushing ever more cash into the AI sector, the ultimate beneficiaries seem to be Broadcom and Marvell, two prominent players in the AI ASIC market. These companies have established themselves as key forces within the landscape, particularly in chip interconnect communication and high-speed data transfer.

Every major player, from Microsoft and Amazon to Google and Meta, as well as generative AI leader OpenAI, is collaborating with Broadcom or Marvell to develop self-designed AI ASIC chips for mass-scale deployment in AI inference computing. As a result, the future market share of AI ASIC is expected to surpass that of AI GPUs, heading toward a relatively balanced market structure, countering the current dominance of AI GPUs which hold approximately 90% of the market.

Recently, Morgan Stanley projected in a report that the AI ASIC market size will expand from $12 billion in 2024 to $30 billion in 2027, with a compound annual growth rate of 34%. However, they assert that the rise of AI ASIC does not necessitate the extinction of Nvidia's AI GPUs; rather, both chip types will coexist, providing complementary solutions for various endpoint demands. Another Wall Street giant, Citibank, indicated that AI ASIC could become increasingly associated with inference tasks, leading to their market share growing in response to the rising demand for inference-level computing.

Additionally, Morgan Stanley utilized a Total Cost of Ownership (TCO) model to compare the cost-effectiveness of AI ASICs and AI GPUs in training and inference tasks. The findings illustrate that ASICs present a lower initial investment, particularly appealing for cloud service providers with budget constraints. Consequently, Morgan Stanley sees positive prospects for Broadcom and Marvell's stock, anticipating they will gain from the surge in inference computing demand spurred by the “DeepSeek wave.”

During earnings calls held by Google and Meta, both Pichai and Zuckerberg signaled an increased commitment to collaborating with chipmakers like Broadcom to develop proprietary AI ASICs. For instance, Google's collaboration with Broadcom resulted in the creation of the TPU (Tensor Processing Unit), which epitomizes an AI ASIC. Previously, Meta designed its first and second-generation processors for AI training/inference acceleration in partnership with Broadcom, with plans to accelerate the development of the next generation AI chip, MTIA 3, by 2025. OpenAI, which recently secured significant investment and collaboration with Microsoft, announced last October plans to develop its first AI ASIC in partnership with Broadcom.

Furthermore, Amazon's management has expressed intentions to scale up the deployment of AI ASIC computing infrastructure, with Marvell being a strategic partner. Last December, Marvell announced a five-year agreement with Amazon AWS to further enhance their AI ASIC strategic collaboration, with plans to produce multiple generations of data center AI chips over the next five years.

Looking ahead, the launch of DeepSeek R1 signifies the dawn of an era characterized by substantial reductions in training and inference costs, marking the arrival of ASICs in the AI arena. Following the unveiling of DeepSeek R1, global tech investors and AI enthusiasts have begun to question the inherent value of Nvidia’s high-performance AI GPUs, leading them to speculate whether collaborations among major players with Broadcom and Marvell for self-designed AI ASICs represent a far more cost-effective solution.

As major model architectures converge towards a few mature paradigms (such as standardized Transformer decoders and diffusion model pipelines), ASICs are better positioned to handle mainstream inference workloads. Furthermore, certain cloud service providers or industry titans may deeply integrate software stacks, enabling ASICs to be compatible with common network operators while providing developers with excellent tools, further accelerating the adoption of ASIC inference in normalized mass-scale scenarios.

Forecasting future computing landscapes, Nvidia’s AI GPUs may increasingly focus on extraordinarily large-scale explorations and rapidly fluctuating multi-modal or novel structural experiments, alongside high-performance computing (HPC), graphic rendering, and visual analytics. Meanwhile, ASICs will harness deep learning-specific operators and data flows for extreme optimization; they prosper in scenarios requiring stable structured inference and high-throughput capacity with high energy efficiency. For instance, if a cloud platform’s AI workloads heavily rely on matrix multiplications, convolutions, layer normalization, and attention mechanisms common in CNNs/Transformers, most AI ASICs will be deeply customized for these operations. AI tasks such as image recognition with ResNet series and Vision Transformers (ViTs), as well as Transformer-based automatic speech recognition or partially fixed multi-modal pipelines, can all benefit from ASIC optimization.

Typically, ASICs employ architectures such as Dataflow or tensor processing units to optimize matrix multiplications, convolutions, activation functions, and attention layers. Once certain large model architectures stabilize in commercial contexts and demand for inference rises sharply, specialized ASIC hardwares that are custom-built can achieve significantly superior energy efficiency and cost-effectiveness relative to general-purpose GPUs (improvements can reach anywhere between 2 to 10 times). Thus, as inference-focused workloads emphasize cost and energy efficiency, ASICs are positioned for broader adoption, particularly for standardizing and scaling AI inference tasks.

As Morgan Stanley posits, both technologies are likely to harmonize in the long term, with a significant market share expansion expected for AI ASICs in the medium term. Nvidia's general-purpose GPUs will concentrate on complex, variable scenarios and cutting-edge research, while ASICs will concentrate on high-frequency, stable, large-scale AI inference workloads and parts of established training processes.