Share with your CTO
Broadcom is positioning itself as the connective tissue of large-scale AI infrastructure, not a headline chip vendor but the company whose networking silicon and custom accelerators keep hyperscaler compute clusters actually running. GraniteShares CEO flagged Broadcom as a top beneficiary of the next AI buildout phase, pointing specifically to power and bandwidth constraints inside data centers as the bottlenecks that shift capital toward Broadcom’s networking and custom silicon portfolio. Analysts forecast earnings growth of 33.73% annually, and the stock trades roughly 12.5% below one fair-value estimate.
What this means for your business
The companies most exposed to this story are the ones writing multi-year infrastructure contracts with hyperscalers right now. If you’re a CTO at an enterprise building on top of cloud AI services, Broadcom’s position matters less to your vendor stack directly and more as a signal: the bottleneck in AI scaling has visibly shifted from model capability to physical infrastructure, meaning bandwidth and power inside data centers are now the constraints that determine what AI workloads can actually run at scale. Organizations already capacity-planning for large inference deployments should be asking their cloud providers pointed questions about interconnect limits.
The infrastructure-before-GPUs thesis deserves scrutiny, though. GraniteShares, as an ETF provider with AI-themed products to market, has an obvious incentive to frame the next wave of AI spending as broad and infrastructure-wide rather than concentrated in Nvidia. That tilt is worth noting, but it doesn’t make the underlying claim wrong. Broadcom’s custom ASIC business with Google (the TPU lineage) and its Tomahawk and Jericho switch silicon genuinely do sit inside every serious AI cluster. The multi-year backlog tied to hyperscalers is real. The risk is customer concentration: a handful of cloud giants represent a disproportionate share of that revenue, and any pullback in their infrastructure spending hits Broadcom harder than it hits a diversified components supplier.
The second-order pressure here lands on procurement cycles, not trading desks. If power and connectivity are the binding constraints on AI cluster performance, then the vendors supplying those layers, Broadcom for silicon, Vertiv and Eaton for power, Arista for software-defined networking, gain pricing leverage in ways that GPU vendors already enjoy but infrastructure vendors historically have not. For a CTO renewing cloud or colocation agreements in the next 12 months, the relevant question is whether your provider has already locked in the networking and power capacity you’ll need, or whether you’re about to compete for constrained infrastructure alongside every other enterprise that delayed that conversation.
Concept deep-dive: Custom silicon (ASICs)
An ASIC, application-specific integrated circuit, is a chip designed to do one job extremely well rather than many jobs adequately, the way a general-purpose CPU does. In AI infrastructure, hyperscalers commission ASICs from companies like Broadcom to handle specific workloads, matrix multiplication for training or high-speed packet routing, at far better power efficiency than an off-the-shelf chip. The business consequence is deep lock-in: once a hyperscaler’s software stack is tuned to a custom chip, switching costs are enormous, which is why Broadcom’s backlog is sticky.
Based on reporting from Broadcom (AVGO) Could Be The Next AI Infrastructure Winner According To GraniteShares, originally published 2026-07-29 21:17:00.
