← Back to feed

Broadcom’s third-quarter earnings, reported after the U.S. market close on September 2, contained both a miss and a moment. The near-term miss: Q4 guidance landed slightly below the most aggressive analyst estimates. The moment: CEO Hock Tan used the earnings call to lay out a revenue trajectory for AI semiconductors that the market had not fully priced in.

The Numbers

Q3 FY2026 results:

• Total revenue: $29.591 billion — up 86% year-over-year, above the revised Wall Street consensus of ~$29.5B
• AI semiconductor revenue: $16.7 billion — up 221% year-over-year, up 54% quarter-over-quarter, above the consensus estimate of $15.9B

Full-year guidance and multi-year outlook:

Period	AI Semiconductor Revenue
FY2026 (ending Oct 2026)	~$58B (raised from $56B during call)
FY2027	~$115B (projected)
FY2028	~$230B (projected)

On earnings per share, Tan projected FY2028 EPS above $30, against a Wall Street consensus that had been continuously revised upward to approximately $26.50.

Marvell Technology, Broadcom’s closest comparable in the custom silicon race, rose 7% on sympathy buying after Broadcom’s report.

What Broadcom Actually Sells

The $16.7B quarter was not from NVIDIA-like GPU compute. It came from two product lines that sit at a different layer of the AI infrastructure stack:

Custom AI ASICs: Broadcom designs and produces application-specific integrated circuits for hyperscalers that want training and inference chips tailored to their specific model architectures. Google’s Tensor Processing Units follow this ASIC route. Amazon, Meta, and others are on similar trajectories. As large model architectures stabilize — fewer novel operators per training run, more repeated inference patterns — custom silicon delivers better cost-per-token and energy efficiency than general-purpose GPUs for the workloads where the pattern is known.

Ethernet switch chips: Every AI cluster above a certain scale needs high-throughput switching between GPU nodes. Broadcom’s networking portfolio has become critical infrastructure for hyperscale AI data centers regardless of which GPU vendor wins the compute layer.

Why the Trajectory Is Credible

The three-year ramp — $58B to $115B to $230B — requires roughly doubling revenue each year for two consecutive years. Aggressive, but grounded in visible customer commitments.

PwC analysis cited during the call showed that internal data center ICT equipment (GPUs, ASICs, storage, high-speed networking) is typically upgraded every four to six years, and that the proportion of capital expenditure going to ICT equipment is rising from approximately 70% today to 93% by 2050. The firm estimates every $1 spent on data center civil construction locks in approximately $12 in subsequent ICT equipment spending over the asset lifecycle.

Hyperscale cloud providers — Google, Amazon, Microsoft, Apple, SpaceX, plus the frontier AI labs they host — have collectively increased capital budgets by approximately 40%, to above $700 billion. That figure is not speculative; it is the sum of disclosed CapEx plans across the five largest operators.

Broadcom has existing multi-year design-win agreements with at least three hyperscalers. Each agreement involves a design cycle measured in years before volume production; the revenue that Tan projected for FY2028 is not speculative demand but contracts already in engineering development today.

The NVIDIA Distinction

Broadcom’s growth does not simply take from NVIDIA. The two address structurally different parts of the same market.

Frontier AI training — pre-training large models with novel architectures — still depends almost entirely on GPU clusters. The operational complexity of adapting a custom ASIC to a rapidly changing model architecture makes general-purpose GPUs the rational choice. NVIDIA’s data center revenue hit $89 billion in its own Q2 FY2027 (reported early September), with guidance for approximately 70% overall revenue growth in FY2028.

What Broadcom captures is the workload that has graduated from the frontier. High-concurrency inference on a fixed model version — the kind of traffic that runs a consumer product at scale — is where custom silicon wins on cost and power. As token invocation volumes across global industries grow exponentially, inference is becoming the larger surface of the market.

The $230B Broadcom projection and NVIDIA’s growth trajectory can coexist. The AI infrastructure buildout is large enough to sustain both, and the compute stack is converging on a heterogeneous architecture: NVIDIA and AMD GPUs for frontier training, custom ASICs for production inference, high-speed Ethernet (also Broadcom) connecting everything.

What to Watch

Broadcom’s Q4 guidance was the one soft note. If hyperscalers pull back on ASIC commitments ahead of NVIDIA’s next-generation product cycle, the FY2027 step-up becomes harder to defend. The company’s exposure to any one hyperscaler — particularly Google, which is the largest ASIC customer by volume — creates concentration risk that the headline revenue projections do not fully surface.

The FY2028 $230B figure assumes continued diversification of the customer base. Broadcom has publicly indicated it is working with additional hyperscale clients beyond its three confirmed design-win relationships. If two or three of those conversations convert to volume commitments before FY2027, the trajectory holds. If they slip, the ramp flattens.