AI inference is moving towards a heterogeneous computing era! Microsoft Corporation's Maia 300 accelerates the rise of self-researched silicon, while Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR (TSM.US) and Marvell (MRVL.US) welcome the ASIC growth cycle.

As media outlets cited sources claiming that one of the American tech giants, Microsoft Corporation (MSFT.US), plans to soon launch its latest self-developed AI accelerator Maia 300, the renowned Wall Street investment firm Wedbush Securities believes that this recent market development could greatly benefit Marvell Technology, Inc. (MRVL.US), which focuses on AI ASICs and data center optical interconnect chips, as well as "the king of chip foundry," Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR (TSM.US). Microsoft Corporation is likely to announce the next generation of its self-developed AI ASIC chip, Maia 300, as early as September 2026, and is in talks with Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR to secure manufacturing capacity for over 300,000 chips by 2027, with a long-term goal of acquiring capacity for over 1 million chips. In a statement, Microsoft Corporation clearly stated that the Maia self-developed AI ASIC deployment ultimately addresses massive AI inference workload demands measured in gigawatts (GW). This largely indicates that hyperscale cloud vendors are making a complete upgrade from "collective GPU purchasing" to a heterogeneous AI computing infrastructure that coexists with NVIDIA Corporation AI GPUs, AMD AI GPUs, and self-developed ASIC/XPUs: AI training operator processes and those complex and rapidly changing cutting-edge AI workloads still heavily rely on AI GPU clusters, while large-scale AI inference workloads around mature and open-source AI models, Copilot, and AI Agents are increasingly suited for dedicated self-developed AI chips. Evidently, Microsoft Corporation is trying to adopt a commercialization path similar to Google TPU of another cloud computing giant, Alphabet Inc. Class C, focusing on power leasing and sales rather than solely using Microsoft Corporations internal AI training or inference processes. However, at present, it resembles cloud power leasing through Azure rather than directly selling bare chips. Microsoft Corporation has explicitly stated that Maia 200 is increasing the AI inference capacity available for actual client use on the Azure cloud computing platform while opening the Maia SDK preview to developers, AI startups, and academic institutions. Overall, Microsoft Corporation's previous generation self-developed AI chip, Maia 200, has transitioned from internal development into the Azure cloud computing production environment and has begun to convert into AI inference capacity that clients can actually consume; however, this should not be interpreted as clients being able to lease Maia 200 chip instances on a large scale and individually like leasing NVIDIA Corporation Blackwell, H100, or Google TPU. More importantly, Reuters recently reported that Microsoft Corporation is attempting to persuade large cloud clients, including Anthropic, to adopt Maia 300, indicating that Maia is no longer just a closed self-developed chip for Microsoft 365 Copilot or internal OpenAI workloads. Microsoft Corporations Maia 300 might impact AI inference costs, while Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR and Marvell stand to be the biggest beneficiaries? Wedbush's senior analyst Matt Bryson wrote in a report to clients: "So far, Microsoft Corporation's Maia self-developed AI chip project has not met expectations, and overall, Microsoft Corporations track record in hardware execution has been lacking." "That said, Microsoft Corporation's self-developed AI chip development work is approaching an important milestone in multi-generational product iteration, and historically, the positive effects brought about by chip development often start to take real progress at this stagesuch as evolving over a span of 4 to 5 years and developing to the third generation of products. If large-scale AI computing infrastructure orders ultimately take shape, the chip IP and design partners of Microsoft Corporations self-developed AI chips, Marvell, and the Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR that manufactures these chips should become direct beneficiaries of Microsoft Corporations self-developed AI chip success." Analyst Bryson also pointed out that Maia 300 may have some impact on NVIDIA Corporation's (NVDA.US) stock price and fundamental outlook, as it will become another custom self-developed AI chip project aimed at reducing enterprise AI inference costs, following Amazon.com, Inc. and Alphabet Inc. Class C's self-developed AI chip systems. However, Bryson added that the impact on NVIDIA Corporation may not be significant because the company has been "very successful at ensuring sufficient supply levels" and in the realm of general-purpose AI computing power. According to media reports, Maia 300 may be released as early as this fall. Microsoft Corporation has been in discussions with Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR to secure manufacturing capacity for over 300,000 chips and plans to deliver by 2027. Maia 200 was launched in January of this year, while early versions of the Maia self-developed AI chip project were first introduced in November 2023. Microsoft Corporation and Marvell did not respond immediately to media requests for comment. Microsoft Corporation aims to replicate the Google TPU path! Maia 300 targets the Azure computing foundation, while Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR and custom chip chains enter a new growth cycle. From an underlying architecture perspective, the advantage that Microsoft Corporation's Maia has over NVIDIA Corporations GPUs is not "absolute computing power," but rather unit economics under ultra-large-scale, repetitive inference loadsthe cost per token, token throughput per watt (Tokens/Watt), and the total cost of ownership (TCO) of AI servers. GPUs must handle training, inference, scientific computing, and many dynamic workloads, so the largest moat is versatility + CUDA/TensorRT software ecosystem + NVLink-level cluster extension capability; Maia, on the other hand, is a customized self-developed AI accelerator designed collaboratively with the workload of Azure, concentrating more transistors, power budgets, and chip area on low-precision tensor computations, memory capacity/bandwidth, and data movement required for Transformers inference. Maia 200 has already adopted Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADRs 3nm process (TSMC 3nm), native FP4/FP8 Tensor Core, 216GB HBM3e, 7TB/s bandwidth, and 272MB on-chip SRAM. Microsoft Corporation has particularly enhanced DMA, NoC, and two-level Scale-up networks, which essentially optimize the architecture to address the increasingly serious memory wall, KV Cache, and data movement costs during the inference stage; Microsoft Corporation's latest financial report indicates that Maia 200 has achieved a 30% improvement in performance per dollar compared to the latest generation of hardware in its fleet, and its self-developed MAI model running on Maia 200 has seen performance improvements of about 40% per watt. This represents the most dangerous competitive edge of ASIC/XPU in the inference era: once billions of similar token generation tasks can be highly standardized, the flexibility premium of NVIDIA Corporations "universal GPUs" might not justify the cost for every inference request. If Maia 300 materializes according to the current media reports in terms of pace and scale, it would be regarded as a "significant structural boon" for the global AI semiconductor investment theme but also means an expansion of the total addressable market (TAM) for AI computing infrastructure and a reshuffling of profit poolsmeaning increased demand for Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADRs advanced process capacity, advanced process supply chain, custom ASIC design, HBM/NAND/data center-level DRAM, advanced packaging, high-speed optical interconnects, and data center networks. However, regarding the expansion expectations for NVIDIA Corporation and AMD-dominated AI GPU technology routes, it may lean toward a negative outlook. Chris Caso, a strategist from the top Wall Street investment firm Wolfe Research, noted that the Philadelphia Semiconductor Index (SOXX) had previously doubled in about three months, then fell roughly 25% from its peak, with recent weakness appearing to be more like an expectation reset after a significant rise. Caso expects demand for AI chips to continue to exceed supply at least until 2028, noting that market concerns about a slowdown in capital expenditure for ultra-large-scale cloud computing have not materialized, and the competitive trend surrounding AI agents is compelling hyperscale cloud vendors "to have no choice but to invest." Senior analysts from Morgan Stanley, led by Brian Nowak, recently published a report indicating a significant upward revision in projected capital expenditures for the five largest ultra-large-scale cloud computing companies (Meta, Amazon.com, Inc., Microsoft Corporation, Alphabet Inc. Class C, SpaceX) for 2027/2028, reaching approximately $1.2 trillion and $1.4 trillion, respectively. The firm also raised its capital expenditure expectations for U.S. tech giants in 2026 from $433 billion a year ago to $805 billion. However, this does not mean that Maia can replace NVIDIA Corporation in the short term. Quite the opposite, the competition in inference has shifted from "chip peak FLOPS" to "overall efficiency of the AI factory," and NVIDIA Corporation's most robust barrier remains its integrated ecosystem of CUDA, TensorRT-LLM, Dynamo, mature kernel libraries, and GPU-NVLink-network-software. TensorRT-LLM has continuously optimized key inference paths such as FP8/FP4 quantization, Paged KV Cache, Speculative Decoding, and Expert Parallelism. Therefore, if Microsoft Corporation truly wants to transform Maia into a large-scale third-party platform, the biggest challenge is not to design the chip but to mature the compiler, kernel, scheduler, model compatibility, and developer ecosystem sufficiently. Institutions like Morgan Stanley and Wedbush Securities, which continue to be optimistic about the investment prospects in the AI computing industry chain, believe that the almost endless demand for cutting-edge computing power and the power demands surrounding AI agents could enable AI ASICs to grow into a second trillion-dollar computing ecosystem without destroying GPU demandfurther reinforcing the investment logic that "the AI semiconductor super cycle is not a single GPU cycle, but a whole data center silicon content enhancement cycle." Concerning the stock price outlook for "the king of chip foundry," Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR, Barclays PLC Sponsored ADR analyst Simon Coles recently raised the target price for Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR from $625 to $650 while maintaining an "overweight" rating; if this price is realized, it means Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR's market value could jump from the current approximately $2.17 trillion to about $3.37 trillion. Coles's core bullish logic can be summarized as follows: the global AI is creating a "battle for advanced logic wafer supply"his Asian supply chain research has already observed strong competition from clients for advanced logic wafer capacity, while Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR concurrently controls the two critical AI computing bottlenecks of N2/N3 advanced processes and CoWoS advanced packaging, and this scarcity is transforming into substantial pricing power. In other words, the bullish scenario of $650 banks on more than just "NVIDIA Corporation continuing to sell GPUs," but also that all heterogeneous AI computing clusters, including NVIDIA Corporation/AMD GPUs + Google TPUs + AWS Trainium + Microsoft Maia + AMD GPUs, ultimately require advanced processes and advanced packaging, positioning Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR as the semiconductor asset most closely approaching the "toll booth" of computing infrastructure in the era of AI inference. Regarding Marvell's stock price outlook, the most aggressive bullish target price currently comes from KeyBanc analyst John Vinh, who raised the target price to $400 from $385 on July 14 and maintained an "overweight" rating; relative to the current trading price of $208.56, this implies an upside potential of approximately 91.8%. Based on an estimated current outstanding share count of about 875.8 million shares, a target price of $400 would imply a potential market capitalization for Marvell of about $350.3 billion. Vinh's core bullish logic revolves tightly around the "heterogeneous computing + AI inference era"that is, Marvell is simultaneously gaining two growth curves from ultra-large-scale cloud vendors for custom AI ASIC/XPU (Custom Silicon) and high-speed interconnect (Optical/Networking). The $400 target price thrown out by KeyBanc analyst Vinh is not simply betting on the Maia 300 chip, but rather on betting that as cloud giants like Microsoft Corporation, Google, and Amazon.com, Inc. transition from "purchasing NVIDIA Corporation GPUs" to the era of heterogeneous computing with GPUs + self-developed ASICs, Marvell will elevate from an AI interconnect supplier to a core "arms dealer" for cloud giants self-developed silicon.