A major shift in AI computing power investment! The market is betting real money on the strong rise of ASIC

According to the Zhitong Finance APP, as of Friday's close of the US stock market, NVIDIA (NVDA.US) shares fell nearly 3%, marking the first significant risk of losing its milestone $4 trillion market capitalization in nearly two months for the world's most valuable company. The "AI chip leader" NVIDIA saw its intraday drop reach nearly 5% on Friday, and from a weekly perspective, NVIDIA's stock price has now fallen for four consecutive weeks. For a stock with a market capitalization exceeding $4 trillion, a nearly 5% drop is considerable, highlighting the market's panic selling triggered by the US non-farm payrolls falling far short of expectations and the pressure from the explosive growth of the Broadcom AI ASIC market.

Additionally, Taiwan Semiconductor's American Depositary Receipts (ADRs) have shown strong gains, making it the second-best performer among recent US chip stocks, only behind Broadcom. In the chip supply chain, Taiwan Semiconductor is regarded as "the eternal god" (YYDS), as both the highly demanded AI GPUs and AI ASICs rely heavily on Taiwan Semiconductor. With decades of chip manufacturing technology accumulation and being at the forefront of chip manufacturing technology improvement and innovation, Taiwan Semiconductor has long dominated the vast majority of global chip foundry orders, especially for the most advanced 5nm and below chip foundry orders.

From a review perspective, the extremely weak non-farm payrolls have cooled expectations for a "soft landing" in the US economy, but the market's expectations for interest rate cuts by the Federal Reserve have surged significantly, with a potential 50 basis point cut in September. This is a positive factor for large-cap growth stocks like NVIDIA. Mark Malek, Chief Investment Officer at Siebert Financial, stated that a 50 basis point move "will bring tailwinds to the stock market." "This will undoubtedly boost large-cap growth stocks and signal investors to take on more risk."

Therefore, the powerful performance and future outlook of Broadcom are at the core of the market's panic logic. NVIDIA's strongest competitor in the AI chip field—Broadcom—has rapidly increased its revenue scale with AI ASICs launched in collaboration with tech giants like Google, Meta, and Microsoft, causing panic among top Wall Street investment institutions and retail investors who have long been optimistic about NVIDIA's stock price and performance expansion trend. Due to the strong revenue scale of AI ASICs, which may soon begin to erode the market share of AI GPUs, they have started to adjust their long-term performance growth expectations for NVIDIA.

AI ASICs and NVIDIA's AI GPUs represent two distinctly different technological routes in AI chips, and currently, they are largely market competitors. Especially for large-scale cloud computing giants and AI leaders like OpenAI, AI ASICs have very clear advantages in cost-effectiveness and energy efficiency in the fields of AI training/inference. The $10 billion "super order" that OpenAI brought to Broadcom and the explosive demand for AI ASICs highlighted in Broadcom's performance have caused Wall Street analysts to see cracks in their performance expectations for NVIDIA from 2026 to 2030. Broadcom's strong rise has forced them to moderately lower the previously high performance growth expectations, which is also the core logic behind NVIDIA's stock price plummeting on Friday

NVIDIA's status in the "4 trillion dollar club" is under pressure - significant stock price pullback in recent weeks.

According to media reports citing informed sources, Broadcom is helping OpenAI design and produce a customized artificial intelligence acceleration chip (i.e., AI ASIC chip). Reports indicate that the two companies plan to deliver the first batch of AI chips from this series as early as next year, which will put NVIDIA, the absolute leader in the AI computing infrastructure field, under the most direct and powerful long-term competitive pressure from Broadcom in the AI chip sector.

Broadcom's semiconductor revenue related to AI infrastructure in the third fiscal quarter was approximately $5.2 billion, with a year-on-year growth rate of 63%, exceeding Wall Street's average expectation of $5.11 billion. Broadcom's management expects this category's revenue to reach about $6.2 billion in the fourth fiscal quarter, indicating a potential year-on-year growth of nearly 70%, higher than analysts' previous expectation of about $5.82 billion. Before the earnings report was released, market expectations for Broadcom's performance and future outlook were very high, so exceeding market expectations significantly boosted investors' bullish sentiment towards Broadcom and the entire AI computing industry chain.

Broadcom's CEO Hock Tan stated during the earnings call that after receiving over $10 billion in AI infrastructure orders from an unnamed new hyperscale customer (later revealed by the media to be OpenAI), the company expects the revenue growth associated with artificial intelligence in fiscal year 2026 to be stronger than previously anticipated. In the last earnings call, Hock Tan mentioned that the AI-related revenue outlook for 2026 would show a growth trajectory similar to this year - with an expected growth rate of about 50% to 60%.

In the Chinese stock market, Cambricon, which is also betting on the ASIC route, has become one of the hottest stocks, with a year-to-date increase of 95%. Goldman Sachs, just a week later, raised its target price for "China's AI chip leader" and "domestic chip replacement" leader Cambricon again. In the latest report released on September 1, Goldman Sachs raised Cambricon's 12-month target price from RMB 1,835 to RMB 2,104, an increase of 14.7%, and maintained a "buy" rating. The latest target price implies a 41% upside from the stock's closing price on August 29, which has been hitting new highs this year.

Recently, Cambricon's stock price and performance have resonated, highlighting the fervent investment boom in AI in China and the "domestic chip replacement" trend. In terms of performance, Cambricon's revenue in the first half of 2025 was RMB 2.881 billion, a staggering year-on-year increase of 4,347.82%, and the company achieved a net profit attributable to shareholders of RMB 1.038 billion, compared to a loss of RMB 530 million in the same period last year The bull market logic of the AI computing power industry chain remains unassailable, but funds are beginning to focus entirely on ASICs

There is no doubt that Broadcom's strong performance and future outlook have reinforced the "long-term bull market narrative" of the AI computing sector. However, the global net long bets are starting to shift significantly from the NVIDIA AI GPU chain to the AI ASIC chain.

The continuously explosive demand for AI computing power globally, coupled with the increasingly large AI infrastructure investment projects led by the U.S. government, and the ongoing massive investments by global tech giants in building large data centers, largely indicate that for long-term investors who are fond of NVIDIA and the AI computing power industry chain, the sweeping "AI faith" around the world has not yet concluded its "super catalysis" on the stock prices of the computing power leaders. They bet that the stock prices of companies in the AI computing power industry chain led by NVIDIA, Taiwan Semiconductor, and Broadcom will continue to exhibit a "bull market curve," thereby driving the global stock market to continue its bull market trend.

It is under the epic price surge of leaders in the AI computing power industry chain such as NVIDIA, Google, Taiwan Semiconductor, and Broadcom, along with the consistently strong performance this year, that an unprecedented AI investment frenzy has swept through the U.S. stock market and global stock markets, propelling the global benchmark index—MSCI World Index to significantly rise since April, recently setting new historical highs.

NVIDIA, with a market capitalization exceeding $4 trillion, saw its intraday drop approach 5% on Friday. In contrast, Broadcom's stock price surged 16% intraday on Friday, reaching a historical high, which increased its market capitalization by nearly $150 billion to about $1.6 trillion, highlighting the global funds flocking to Broadcom and the rapidly warming bullish sentiment towards this AI ASIC leader.

Jim Awad, Senior Managing Director at Clearstead Advisors, stated that while investors should prepare for NVIDIA to face stronger competition in this field, the rapid growth of the AI infrastructure market means that even if the company loses some market share, it can still maintain a continuous growth trend. However, its future growth rate may continue to lag behind competitors like Broadcom.

With this ongoing decline, NVIDIA has dropped about 10% from its August peak, with a market value evaporating by nearly $470 billion. The company recently fell below its 50-day moving average; nevertheless, it remains the highest market capitalization company globally. Microsoft ranks second with a market capitalization of $3.7 trillion.

Daniel O’Regan, Managing Director of Equity Trading at Mizuho Securities, pointed out that compared to Broadcom, NVIDIA has reached a relative low not seen in 18 months, which is quite noteworthy. “The term OpenAI indeed generates a lot of momentum, and this accelerates a trend: funds are increasingly favoring Broadcom,” he stated, noting that Broadcom's stock performance has significantly outperformed NVIDIA this year.

“Today, the divergence between the two seems somewhat extreme, but NVIDIA has been synonymous with AI for about the past three years. While I don’t think the market is cooling on it, funds are indeed expanding to other AI winners,” O’Regan emphasized. “In contrast, Broadcom is a ‘shiny new thing.’” On Wall Street, given the continuous explosive growth in demand for Broadcom's Ethernet switch chips and AI ASIC chips, Wall Street financial giants are generally bullish on Broadcom's stock price outlook, optimistic that Broadcom will continue to set new highs in its stock price. As a result, after Broadcom announced its earnings, many analysts significantly raised their target price for the company over the next 12 months. Susquehanna, Bernstein, KeyBanc, and Barclays have all significantly raised their target price for Broadcom to $400. As of Friday's market close, Broadcom's stock rose nearly 10% to $334.89.

The AI ASIC wave is becoming increasingly fierce

With its absolute technological leadership in inter-chip communication and high-speed data transmission between chips, Broadcom has become one of the most important players in the AI infrastructure field for customized AI ASIC chips in recent years. For example, in Google's data center server AI chip—TPU AI acceleration chip, Broadcom is a core participant, collaborating with Google's team in the development of the TPU AI acceleration chip. In addition to chip design, Broadcom also provides Google with critical inter-chip communication intellectual property and is responsible for manufacturing, testing, and packaging new chips, thereby safeguarding Google's expansion into new AI data centers.

As American tech giants firmly invest heavily in the field of artificial intelligence, the biggest beneficiaries include not only NVIDIA but also AI ASIC giants like Broadcom, Marvell Technology, and Taiwan's Wistron. Microsoft, Amazon, Google, and Meta, as well as generative AI leader OpenAI, are all working with Broadcom or other ASIC giants to update and iterate AI ASIC chips, primarily used for massive inference AI computing power deployment. Therefore, the future market share expansion of AI ASICs is expected to significantly outperform AI GPUs, moving towards equal shares rather than the current situation where NVIDIA's AI GPUs dominate—holding up to 90% of the AI chip market.

It is understood that one of Broadcom's major clients, Google, disclosed the latest details of the Ironwood TPU (TPU v6) at the conference, showcasing remarkable performance improvements. Compared to TPU v5p, Ironwood's peak FLOPS performance has increased tenfold, with an efficiency ratio improvement of 5.6 times, and compared to Google's TPU v4 launched in 2022, Ironwood's single-chip computing power has increased by more than 16 times.

Performance comparisons show that Google's Ironwood has a power efficiency of 4.2 TFLOPS/watt, which is slightly lower than NVIDIA's B200/300 GPU's 4.5 TFLOPS/watt. JP Morgan commented that this performance data highlights that advanced AI's dedicated AI ASIC chips are rapidly narrowing the performance gap with market-leading AI GPUs, driving hyperscale cloud service providers to increase investments in more cost-effective customized ASIC projects According to the latest forecast from Wall Street financial giant JP Morgan, the chip, developed in collaboration with Broadcom using advanced 3nm process technology, is expected to be mass-produced in the second half of 2025, and Ironwood is projected to bring approximately $10 billion in revenue to Broadcom over the next 6-7 months.

The cost-performance advantage of AI ASICs represents the biggest advantage for giants like Google, Microsoft, and OpenAI compared to NVIDIA's AI GPUs.

The MTIA v2 AI ASIC within Meta's data centers has shown to consume less energy than NVIDIA's H100 under third-party GEMM/inference testing models, while achieving similar throughput (both approximately 5.5–5.7× relative to T4). If we estimate the maximum costs at H100 ≥ $30,000 and MTIA v2 at $2,000–$3,000, the cost-performance metric (perf/$) of MTIA is far superior to that of NVIDIA.

Therefore, while AI ASICs may not completely and massively replace NVIDIA, their market share is bound to expand, moving away from the current situation where NVIDIA's AI GPUs dominate. Especially in the actual configuration of AI data center computing infrastructure, the "mixed formation" of AI ASICs and NVIDIA AI GPUs (GPUs for training/exploration, ASICs for large-scale inference/partial training) significantly enhances energy efficiency and greatly reduces total cost of ownership (TCO).

In standardized mainstream inference and partial training (especially for continuous long-tail training/fine-tuning), the "unit throughput cost/energy consumption" of customized AI ASICs is significantly better than pure GPU solutions; however, for rapid exploration, cutting-edge large model training, and multimodal new operator experimentation, NVIDIA's AI GPUs remain the mainstay. Therefore, in current AI engineering practices, tech giants are increasingly inclined to adopt a hybrid architecture of "ASICs for normalization and GPUs for peak exploration/new model development" to minimize TCO.

Taiwan Semiconductor exemplifies what it means to be "timeless"!

The booming demand for AI GPUs since 2023, along with the recent explosive demand for AI ASICs, is closely tied to Taiwan Semiconductor. TSMC is currently the world's largest contract chip manufacturer, and as the fervor for AI continues unabated and sweeps across the globe, its clients, including NVIDIA, AMD, and Broadcom, continue to benefit from the surge in demand for the most essential infrastructure for AI—AI chips. The scale of chip manufacturing contracts from these chip giants to TSMC has surged, driving TSMC's performance to consistently exceed expectations since last year, which is a key logical support for the continuous record highs in TSMC's stock prices in Taiwan and its ADRs in the U.S. this year.

With decades of chip manufacturing technology accumulation and being at the forefront of chip manufacturing technology improvement and innovation (pioneering the FinFET era and driving the start of the 2nm GAA era), TSMC has maintained a dominant position in the global chip manufacturing market with advanced processes and packaging technologies, as well as ultra-high yield, long dominating the vast majority of global chip foundry orders, especially for 5nm and below advanced process chip foundry orders More importantly, TSMC is currently capturing almost all high-end chip packaging orders at 5nm and below with its industry-leading 2.5D and 3D chiplet advanced packaging, and the advanced packaging capacity is far from meeting demand. NVIDIA's Blackwell has been in short supply since it achieved mass production at the end of last year, primarily constrained by TSMC's 2.5D CoWoS packaging capacity. Currently, chip giants like Apple, AMD, NVIDIA, and Broadcom are turning to TSMC's 3D advanced packaging capacity, which may further drive the supply-demand imbalance in TSMC's advanced packaging capacity.

TSMC's latest performance shows that the demand for AI computing power has surged, driving TSMC's net profit in Q2 to increase by 61%. TSMC expects that sales in 2025, calculated in US dollars, will grow by about 30%, higher than the previous expectation of "close to 20% in the mid-range," mainly due to the continuous surge in AI chip orders based on 3nm and 5nm advanced process technology. As the demand for AI computing power remains incredibly strong, TSMC is actively expanding its backend capacity to enhance the actual output of CoWoS advanced packaging, primarily for NVIDIA's AI GPU capacity, which also indicates the company's confidence that the strong demand for AI chips will continue until 2026.