Short-term growth fades, but the long-term AI computing power blueprint shines! Major clients ignite the ASIC super cycle, and Broadcom (AVGO.US) stuns with a $230 billion AI semiconductor outlook

According to the Zhitong Finance APP, one of the biggest winners in the global AI boom, Broadcom (AVGO.US), released its financial results for the third quarter of fiscal year 2026 and the latest management outlook after the US stock market closed on September 2 (morning of Thursday, Beijing time). The results show that Broadcom's Q3 revenue was $29.591 billion, up 86% year-on-year, exceeding the strong upwardly revised Wall Street consensus estimate of about $29.5 billion. The company's total AI semiconductor revenue—including AI ASIC (Google's TPU belongs to this technical route) and Ethernet switch chips—reached $16.7 billion, a 221% YoY increase and a 54% QoQ increase, surpassing the consensus estimate of $15.9 billion.

For its highly anticipated future outlook, management further projects that AI semiconductor revenue will expand massively from roughly $58 billion in fiscal 2026 (an upward revision from $56 billion previously) to $115 billion in fiscal 2027, and potentially a staggering $230 billion in fiscal 2028. The company also expects earnings per share to exceed $30 in fiscal 2028—significantly higher than Wall Street's upwardly revised consensus of approximately $26.5.

Broadcom is a core chip supplier to Apple and other major tech firms, as well as a leading provider of high-performance Ethernet switch chips for global large-scale AI data centers and a major custom AI ASIC chip supplier, which is crucial for the cloud giants’ self-developed AI chips for AI training/inference.

Broadcom's extremely strong results and outlook highlight, with the arrival of the AI inference era and surging demand for cloud AI inference computing power—as well as the trend of "micro-training" embedding AI models into enterprise operations—that higher cost-performance and energy-efficient AI ASIC systems are posing a strong challenge to Nvidia’s nearly 90% monopoly in the AI chip market. Broadcom’s results and strong outlook, together with Nvidia’s explosive earnings and approximately 70% growth expectation for overall revenue in fiscal 2028, underscore that global artificial intelligence computing demand remains in a massive expansion cycle that is far from its peak.

Global AI application leaders such as OpenAI and Anthropic, along with hyperscale cloud computation providers such as Google, Amazon, and SpaceX, are advancing the construction of AI computing infrastructure from “joint purchases of Nvidia GPUs” to a fully upgraded heterogeneous computing system of Nvidia GPU + AMD GPU + custom AI ASIC/XPU + CPU/DPU in large-scale collaborative operation. As AI model frameworks gradually stabilize and token invocation across industries grows exponentially, high-concurrency AI inference workloads are increasingly suitable for lowering unit token costs using custom ASIC chips.

In other words, processes for AI model training and the most complex, rapidly evolving cutting-edge AI workloads still require AI GPU clusters—frontier pre-training, reinforcement learning, and new rapidly changing operators still rely more on the programmability, CUDA ecosystem, and NVLink/NVSwitch cluster capabilities of GPUs. Meanwhile, large-scale AI inference workloads around mature and open-source AI models, Copilot agent-style AI workflows, and AI Agent operations are becoming increasingly suitable for dedicated custom AI ASIC chips.

According to optimistic institutions like Morgan Stanley and Wedbush Securities, the virtually endless cutting-edge computing power and demand for AI Agent-centric processing in the AI inference era allows AI ASIC to become a second trillion-dollar-level ecosystem—without cannibalizing GPU demand—thereby enhancing the logic that “the AI semiconductor supercycle is not merely a GPU cycle, but a supercycle of overall data center silicon content growth.”

Short-term guidance slows down, long-term computing power blueprint is stunning: Broadcom aims for $230 billion AI semiconductor revenue in two years, challenging Nvidia's GPU dominance

Broadcom’s Q3 FY2026 revenue was $29.591 billion, up 86% YoY, above the Wall Street mean estimate of about $29.5 billion; adjusted EPS was $3.32, a 96% YoY increase and higher than the $3.23 consensus; total AI semiconductor revenue reached $16.7 billion, up 221% YoY and 54% QoQ, exceeding the $15.9 billion consensus.

Broadcom’s management gave Q4 revenue guidance of $34.8 billion, a 93% YoY increase, slightly below the Wall Street consensus of $35.1 billion; this was the most direct trigger for Broadcom’s stock pressure in after-market US trading after earnings, with the stock falling more than 6% post-market, then bouncing nearly 3% and continuing to fluctuate. However, guidance for Q4 AI semiconductor revenue reached $21.7 billion, a 236% YoY jump, exceeding the consensus of around $21.3 billion. On the call, management further projected AI chip revenue to climb from about $58 billion in FY2026 to $115 billion in FY2027 and $230 billion in FY2028, with EPS over $30 in fiscal 2028.

The market's requirement for Broadcom’s growth and outlook now is not simply “grows fast” but “can it continually and substantially beat expectations?” Before earnings, the stock had fallen more than 20% from its early June peak, with market cap evaporating by over $520 billion, even though year-to-date as of June 2, the stock had soared by 50%. Given the valuation system for leading AI chip companies, investors are hoping management will provide long-term, AI compute-related revenue visibility similar to Nvidia last week, and replicate strong Nvidia-style guidance.

This quarter Broadcom only provided routine guidance for Q4 revenue, with total guidance slightly below expectations and without meeting the market’s desire for longer-term, quantifiable revenue paths. In the thin after-hours market, this amplified selling pressure.

Broadcom’s core logic for strong growth is not that ASIC will immediately replace Nvidia AI GPUs, but that hyperscalers are building heterogeneous architectures of “general purpose GPU + custom XPU”: GPUs for general training and rapid iteration, ASICs like TPUs for performance, cost, energy efficiency, and supply chain autonomy for stable and internal workloads. With expertise in custom AI accelerators, high-speed data center interconnects, SerDes, switch chips, and Ethernet infrastructure, Broadcom benefits both from increased chip shipments and the rising value of high-speed interconnects as cluster complexity and demand expand.

In Q3, GAAP operating profit was $15.955 billion, up 171% YoY; GAAP net profit was $13.088 billion, up 216%; GAAP diluted EPS $2.68, up 215%. Non-GAAP operating income was $20.095 billion, up 92%; non-GAAP net profit $16.372 billion, up 95%. Operating cash flow was $14.197 billion, up 98%; after $532 million in capex, free cash flow was $13.665 billion, up 95%, with a free cash flow margin of 46%.

Q3 semiconductor solutions revenue was $20.839 billion, up 127% YoY, accounting for 70% of total revenue; infrastructure software revenue was $8.752 billion, up 29%, accounting for 30%. Quarter-end cash and equivalents stood at $24 billion, up from $19.6 billion last quarter. Non-GAAP gross margin was about 75%, operating margin about 67.9%. Despite higher HBM memory costs in AI accelerators suppressing the margin percentage, scale and operating leverage still expanded absolute profit.

According to the call, Q3 XPU shipments grew over 3.5 times YoY, contributing about 73% of AI semiconductor revenue; non-AI semi revenue was about $4.2 billion, up 5% YoY, essentially flat QoQ. Q4 semiconductor revenue is expected to be about $26.1 billion, up 136% YoY; management expects Q4 infrastructure software revenue of about $8.7 billion, up 24–25% YoY; non-GAAP gross margin projected to decline to about 73%, capex to rise to $1.4 billion.

As Alphabet’s Google, OpenAI, Meta Platforms, and other companies seek to diversify suppliers beyond Nvidia, Broadcom’s custom AI semiconductor business is booming.

Chatbot developers racing to build AI infrastructure, such as Anthropic and OpenAI, have become especially important Broadcom customers. By 2027, Anthropic is expected to overtake Google as Broadcom’s largest custom AI ASIC customer, and CEO Hock Tan projected on the earnings call that OpenAI will soon become the second largest customer.

Rapid AI data center buildout is also boosting Broadcom’s networking product sales. Bloomberg Intelligence research analysts Kunjan Sobhani and Oscar Hernandez Tejeda noted in a report that the five largest hyperscale cloud service providers—as major data center operators—have increased capital budgets by about 40% to over $700 billion, “significantly improving demand visibility for Broadcom’s custom AI chip and networking business.”

This long-term prospect has excited investors; by contrast, Broadcom’s Q4 outlook did not impress as much. The company said the quarter ending October will bring in $34.8 billion. Compiled data shows Wall Street analysts, on average, expected about $35.1 billion, with some projecting well over $36 billion. Broadcom stated AI chips alone will bring in $21.7 billion in Q4, slightly above expectations, although some estimates far exceeded $22 billion.

At the close of regular trading, Broadcom shares are up 6.1% year-to-date, lagging many of its peers’ gains in 2026.

On the call, CEO Hock Tan signaled accelerated cooperation with Google, Anthropic, and OpenAI, saying the company will deliver “tens of billions of dollars” in custom processors annually to Google in the coming years. Using energy consumption (gigawatts) as a proxy for data center capacity, Broadcom will ship chips supporting 5 GW next year to Anthropic, and another 10 GW the following year.

Tan also stated that Broadcom “can foresee” delivering more than 5 GW worth of custom chips to OpenAI in 2028. Between now and the end of 2027, Broadcom will supply three generations of chips to its fourth-largest customer, Meta.

Hock Tan has always positioned Broadcom as a strong alternative to Nvidia’s leading chips, which currently dominate the AI field. On the call, he noted somewhat provocatively that Broadcom and Google’s next-generation chip for AI models, even if not outperforming Nvidia’s new-generation Vera Rubin lineup, at least matches it.

Broadcom has benefited from surging custom AI chip demand—not only continuously winning deals, but also pioneering funding tools, for example with Apollo Global Management and Blackstone, to help Anthropic finance expensively sourced semiconductors designed with Broadcom and Google. This financing partnership aims to support over 20 GW of computing capacity, requiring hundreds of billions in funding—roughly the output of 20 nuclear power plants.

Nvidia's 70% growth outlook and Broadcom’s strong guidance shatter “AI peaking” arguments, perfectly demonstrating that global AI computing infrastructure build-out remains in full swing

The Philadelphia Semiconductor Index’s strong rebound, Nvidia’s recent results and 70% growth outlook, Broadcom’s robust results and guidance, Anthropic’s newly signed multi-billion dollar AI cloud compute contracts, and Korea’s surging semiconductor export data all highlight that global demand for AI computing power remains in a significant upward supercycle.

After the July AI deleveraging and unwinding pullback, investors remain strongly risk-on regarding the AI infrastructure chain. As of September 2, the Philadelphia Semiconductor Index was up 60.09% year-to-date, and the Korea KOSPI Index rose 55.73%. Both briefly slipped into technical bear markets in July (falling by 20%), then rebounded sharply—KOSPI, often seen as a barometer for global AI compute investment, rebounded over 30% from its July lows into a new technical bull. As of September 1, the Philadelphia Semiconductor Index closed at 11,339.25, up about 60% YTD; after a nearly 29% drawdown from June’s high to July 29, it crossed the technical bull line with a 20% rally from the recent bottom.

AI infrastructure demand is becoming long-term locked-in compute capacity, not just cloud budget intent. Anthropic, for example, reportedly signed a $35 billion (about 350 MW capacity) AI cloud deal with Nvidia-backed Lambda; recently it locked in about $45 billion for 460 MW of capacity over six years with Nscale, which will deploy Nvidia’s next-gen Vera Rubin platform at scale. Korea’s August exports surged 68.7% YoY to $98.26 billion; semiconductor exports soared 209% to a record $46.65 billion, accounting for 47.5% of total exports.

AI GPU clusters, AI ASIC (TPU) clusters, and storage chips remain the bottleneck in global AI infrastructure under the token inference tidal wave. Counterpoint expects AI server AI ASIC shipments to triple by 2027, with Broadcom potentially reaching a 60% market share among ASIC design partners; TrendForce projects Q3 DRAM and NAND Flash contract prices to rise 13–18% and 10–15%, respectively. Through 2026, server DRAM and enterprise SSD prices are expected to cumulatively rise about 270% and 235%, while 2027 HBM contract prices could rise another 70–140%. DRAM and NAND are forecast to comprise 47% of cloud capex, rising to 68% by 2027.

Of 33 Wall Street analysts surveyed by MarketBeat, 29 rate Broadcom as a “Buy,” 4 as “Hold,” none as “Sell”; the average target price is $491.97, implying 33.96% upside potential, while TIPRANKS reports an even more optimistic $509 target. The general outlook is bullish, but weaker Q4 guidance, lower AI product gross margins, Google introducing second suppliers, and customer and financing risks remain factors influencing near-term valuation swings.

Citi lists Broadcom as its top semiconductor pick; Morgan Stanley expects Broadcom to maintain about an 80% serviceable ASIC design market share; Deutsche Bank, Goldman Sachs, and Bank of America agree Broadcom’s growth cycle is being extended by custom AI ASIC/XPU clusters, AI Ethernet networking infrastructure, and customer diversification—with BofA setting a target price as high as $530.

Nvidia’s Q2 revenue was $96.2 billion, especially its $89 billion data center division revenue, up 106% and 117% respectively; the 70% growth outlook for fiscal 2028 is significantly higher than Wall Street’s previous 44% expectation. Morgan Stanley recently noted that, given abundant wafer and HBM supply, Nvidia’s growth could exceed 100%. All of this shows that Nvidia GPUs and Broadcom’s custom AI accelerators (AI ASIC/TPU/Custom Accelerator/XPU) are accelerating simultaneously—not simply shifting market share but rapidly expanding the overall AI computing pool.

PwC’s latest estimates show this is not a one-off, property-style capital expenditure but a recurring “AI compute subscription cycle”: in its base case, PwC projects total global data center investment of a staggering $31.6 trillion from 2026–2050, and nearly $50 trillion in a most optimistic AI adoption scenario; annual spending rising from about $800 billion in 2026 to $1.1 trillion in 2030 and $1.8 trillion in 2050. The US will absorb $15.1 trillion (48%), Asia-Pacific $8.2 trillion.

PwC estimates also show that AI GPU, AI ASIC/TPU, storage, and high-speed networking upgrades inside data centers occur every 4–6 years, increasing the share of ICT equipment in data center capex from about 70% today to 93% by 2050. PwC also estimates that for every $1 in data center construction capex, about $12 in subsequent ICT investment will be driven over the asset’s lifetime—i.e., every $1 in physical build will trigger about $12 more for internal upgrades such as AI GPU, AI ASIC/TPU, and storage chips.