Broadcom (AVGO.US) posted fiscal Q3 2026 earnings on Thursday morning Beijing time, revealing a dramatic surge in AI-driven revenue but offering a near-term outlook that briefly spooked investors. The company's total revenue hit $29.591 billion, up 86% year-over-year and edging past the elevated consensus estimate of roughly $29.5 billion. AI semiconductor revenue, which encompasses custom AI ASICs like Google's TPU and Ethernet switch chips, soared 221% to $16.7 billion, exceeding the $15.9 billion expectation. However, the real headline was the company's multi-year blueprint for AI dominance. Management now projects AI semiconductor revenue will explode from approximately $58 billion in fiscal 2026 to $115 billion in fiscal 2027, and a staggering $230 billion by fiscal 2028. That outlook was coupled with a forecast for fiscal 2028 EPS above $30, significantly topping Wall Street's upwardly revised consensus of roughly $26.50.

As a key chip supplier to Apple and other tech titans, plus a dominant force in high-performance Ethernet switch chips for data centers, Broadcom is at the heart of the AI infrastructure build-out. The company’s robust results and forward guidance underscore a pivotal shift. As the AI landscape moves toward inference-heavy workloads and "micro-training" for enterprise-specific models, the superior cost-efficiency and performance-per-watt of custom ASIC systems are mounting a serious challenge to Nvidia's near-90% grip on the AI chip market. This news, combined with Nvidia's own blockbuster earnings and its projection of a 70% revenue surge in fiscal 2028, collectively signals an AI compute demand cycle that is nowhere near its peak.

The AI infrastructure buildout, led by hyperscalers like Google, Amazon, and OpenAI backer Microsoft, is evolving from a simple "buy Nvidia GPUs" approach to a complex heterogeneous computing fabric. This new architecture runs Nvidia and AMD GPUs alongside custom AI ASICs (XPUs), CPUs, and DPUs in a coordinated fashion. As AI model architectures stabilize and token usage across industries grows exponentially, the high-concurrency workloads of AI inference are increasingly better served by custom ASICs to drive down unit costs. While the most complex frontier training jobs will still rely on the programmability and ecosystem of Nvidia's GPU clusters, the massive scale of inference for mature models, copilots, and AI agents is creating a massive new market for specialized silicon. Investment banks like Morgan Stanley and Wedbush view this as a second trillion-dollar compute ecosystem emerging, not by destroying GPU demand, but by expanding the overall silicon content within data centers, reinforcing the narrative of a super-cycle for all AI compute components.

For the fourth quarter, Broadcom guided total revenue to $34.8 billion, a 93% year-over-year jump but slightly below the Wall Street average of $35.1 billion. This slight shortfall triggered a volatile after-hours response, with shares initially sliding over 6% before recovering. The more specific AI semiconductor guidance for Q4 was $21.7 billion, a 236% increase and slightly above expectations of $21.3 billion. On the conference call, CEO Hock Tan reiterated the explosive growth trajectory for AI chips, moving from ~$58 billion in 2026 to $115 billion in 2027 and $230 billion in 2028, with EPS surpassing $30 in that final year. The market's reaction highlighted that for a stock of Broadcom's stature, growth is no longer enough; investors demand a beat-and-raise. Having already fallen over 20% from its early June peak, wiping out over $520 billion in market value, the stock's year-to-date gain had been trimmed to just over 6% ahead of the earnings release. The market hunger for a long-term revenue path was only partially satisfied, and the slightly softer Q4 total revenue guide provided a focal point for selling pressure in illiquid after-hours trading.

The core thesis for Broadcom's growth isn't about ASICs immediately replacing Nvidia GPUs, but rather about hyperscalers building a "general-purpose GPU + custom XPU" infrastructure. GPUs remain the workhorse for general training and rapid iteration, while ASICs like TPUs optimize power efficiency, cost per token, and supply chain autonomy for stable training and massive inference workloads. Broadcom’s unique position spans custom accelerators, high-speed SerDes, switching chips, and Ethernet networking, allowing it to benefit from both increasing chip counts and the rising complexity and value of the interconnect fabric within expanding clusters. The fiscal Q3 reported results showed GAAP operating income of $15.955 billion (up 171%), and GAAP net income of $13.088 billion (up 216%). On a non-GAAP basis, operating income was $20.095 billion (up 92%), net income was $16.372 billion (up 95%), and free cash flow reached $13.665 billion, maintaining a robust 46% margin. Segment-wise, semiconductor solutions revenue grew 127% to $20.839 billion, while infrastructure software contributed $8.752 billion, up 29%. The non-GAAP gross margin held at ~75%, with an operating margin of ~67.9%.

During the call, management revealed that XPU shipments more than tripled year-over-year and now account for ~73% of AI semiconductor revenue. Non-AI semiconductor revenue was flat sequentially at ~$4.2 billion. For Q4, they guide semiconductor revenue to ~$26.1 billion (up 136%) and infrastructure software to ~$8.7 billion (up ~24-25%). The non-GAAP gross margin is expected to dip to ~73% due to the higher cost content of AI accelerators and HBM memory, while capital expenditures are projected to rise to $1.4 billion. The business is booming as customers like Google, OpenAI, and Meta look to diversify their AI chip supply chains. Anthropic and OpenAI, both racing to build massive AI infrastructure, have become crucial clients. CEO Hock Tan expects Anthropic to replace Google as Broadcom's largest custom ASIC customer by 2027, with OpenAI becoming the second-largest, and has already outlined massive delivery commitments. Analysts at Bloomberg Intelligence noted that the combined capex budgets of the top five hyperscalers have surged ~40% to over $700 billion, providing exceptional visibility for Broadcom's custom silicon and networking businesses.

While the long-term story excited investors, the near-term Q4 total revenue forecast of $34.8 billion fell short of the $35.1 billion average estimate and some forecasts above $36 billion. However, the AI-only chip revenue guide of $21.7 billion for Q4 was slightly ahead of the $21.3 billion average. Hock Tan detailed deepening partnerships, stating Broadcom would ship "tens of billions" of dollars worth of custom processors to Google annually, and would deliver chips to Anthropic supporting 5 gigawatts of compute capacity next year, followed by another 10 gigawatts the year after. He also "sees a path" to provide OpenAI with more than 5 gigawatts of custom chips by 2028, and will deliver three generations of chips to Meta by the end of 2027. Tan has been positioning Broadcom as a potent alternative to Nvidia, even suggesting that a new chip co-developed with Google rivals or exceeds the performance of Nvidia's next-gen Vera Rubin platform. Beyond chip design, Broadcom has pioneered financing solutions, setting up vehicles with Apollo Global Management and Blackstone to help Anthropic fund its massive semiconductor procurement, supporting a compute capacity exceeding 20 gigawatts—a scale that requires hundreds of billions of dollars and rivals the output of 20 nuclear power plants.

The powerful combination of Nvidia's 70% future growth outlook and Broadcom's stellar guidance delivers a decisive blow to the "AI peak" narrative. This is reinforced by Anthropic's recent multi-billion dollar cloud compute agreements and South Korea's record semiconductor exports. The Philadelphia Semiconductor Index (SOX) remains up over 60% year-to-date, and South Korea's KOSPI, a bellwether for AI infrastructure, has rebounded over 30% from its July low into a new technical bull market. These positive signals across the AI supply chain affirm that global demand for AI compute power is still in a significant upward cycle. Demand is now being converted from budget intentions into multi-year capacity reservations, with Anthropic securing hundreds of megawatts of compute in recent deals. South Korea's August exports surged, with semiconductor shipments skyrocketing 209% to a record $46.65 billion, comprising 47.5% of total exports. Industry forecasts point to a tripling of AI server ASIC shipments from 2024 to 2027, with Broadcom poised to hold roughly 60% of the ASIC design partner market in that period. Concurrently, memory prices are expected to rise sharply, with server DRAM and enterprise SSD prices potentially climbing 270% and 235% respectively in 2026, as the cost of these components becomes an ever-larger portion of cloud service provider capex.

Wall Street's response is generally bullish. Of 33 analysts tracked by MarketBeat, 29 rate Broadcom a "Buy" with an average price target of $491.97, implying a 33.96% upside. Citi lists Broadcom as a top semiconductor pick, Morgan Stanley expects it to retain ~80% of the serviceable ASIC design market, and Bank of America is notably bullish with a $530 target. The collective thesis from Deutsche Bank, Goldman Sachs, and BofA is that custom AI ASIC/XPU clusters, AI Ethernet networking, and customer diversification are extending Broadcom's growth cycle. This is set against Nvidia's own staggering performance, where its data center revenue alone hit $89 billion. JPMorgan suggests Nvidia's growth could even exceed 100% if supply were fully unconstrained. The concurrent acceleration of both Nvidia's GPUs and Broadcom's custom accelerators proves the overall AI compute demand pool is expanding intensely, not merely shifting market share. PwC's recent analysis frames this as an "AI compute subscription cycle," projecting cumulative global data center investment of $31.6 trillion from 2026 to 2050, potentially approaching $50 trillion in a faster-adoption scenario. Annual spending is expected to rise from about $800 billion in 2026 to $1.1 trillion by 2030, with the US absorbing ~48% of cumulative investment. The consulting firm highlights that internal ICT equipment, such as AI GPUs, ASICs, storage, and networking, is refreshed every four to six years, meaning it will account for 93% of data center capex by 2050, and estimates that every $1 invested in data center construction locks in approximately $12 of future ICT equipment spending.