Broadcom’s stock is rising. Here’s why its new Google and Anthropic deals are so significant.
The short answer
Investor enthusiasm rests on one idea: Broadcom is becoming an indispensable supplier for the AI compute build‑out. Fresh or expanded engagements with Google and Anthropic reinforce Broadcom’s position across three profit pools—custom AI silicon, high‑speed networking, and advanced packaging—while improving demand visibility and deepening customer lock‑in.
What the Google tie‑up signals
– Deeper custom silicon role: Google’s TPU program has long relied on Broadcom for co‑design, networking, and physical implementation. A renewed or expanded deal indicates Broadcom remains in the loop for future TPU generations and adjacent infrastructure (switch silicon, optical connectivity, and possibly package integration).
– End‑to‑end stack leverage: Google’s AI datacenters depend on more than accelerators. Broadcom supplies Ethernet switch ASICs (Tomahawk/Jericho families), PCIe/CXL connectivity, and optical PHYs—components that scale in tandem with every new AI pod Google deploys.
– Advanced packaging capacity: The hardest bottleneck in AI hardware is not just wafers; it’s advanced packaging (e.g., CoWoS‑class integration). Being Google’s partner helps Broadcom secure scarce packaging capacity and sharpen its know‑how, a durable advantage as chiplets and HBM stacks proliferate.
Why an Anthropic deal matters even more than the logo
– Proof that model labs want custom chips: If Anthropic is co‑developing accelerators with Broadcom, it validates a broader trend—foundation‑model companies seeking silicon tuned to their training and inference workloads to cut cost per token and reduce exposure to GPU shortages.
– A new customer class: Hyperscalers like Google are one profit stream; model companies are another. Winning a leading AI lab demonstrates Broadcom’s semi‑custom model travels beyond cloud titans.
– Design wins compound: Even if initial volumes start modestly, a working custom chip can scale rapidly as model sizes, context windows, and inference endpoints grow. Each successive tape‑out benefits from prior software, kernels, and packaging IP.
The technology edge Broadcom brings
– Custom accelerators: Broadcom’s ASIC organization co‑designs cores, memory subsystems, and interconnects, then shepherds chips through physical design and bring‑up. NRE dollars today can lead to multi‑year production revenue once a design ramps.
– AI networking at massive scale: Broadcom’s merchant Ethernet (Tomahawk) and AI‑optimized fabric (Jericho‑AI) target ultra‑large clusters with RoCE and congestion control tuned for GPU/accelerator traffic. As AI training shifts toward Ethernet in some deployments, Broadcom benefits regardless of which accelerator wins.
– Optics and PHYs: Faster links (800G today, moving toward 1.6T) need PAM4 DSPs and retimers. Broadcom’s optical components ride every bandwidth upgrade.
– Advanced packaging: Integrating HBM stacks, chiplets, and high‑power dies demands packaging expertise and access to capacity. This capability is a strategic chokepoint and a margin lever.
Financial implications investors like
– Better revenue visibility: Multi‑year custom engagements come with roadmaps and volume commitments, smoothing the order book versus pure merchant sales.
– Top‑line acceleration with operating leverage: Each AI deployment lifts multiple Broadcom products—custom silicon, switches, optics, connectivity—creating a multiplier effect per dollar of customer capex.
– Attractive margin mix over time: While custom ASICs can start with lower gross margins than mature merchant chips or software, scale, NRE amortization, packaging services, and high‑value IP typically lift margins as programs ramp.
– Optionality beyond one platform: Broadcom doesn’t need to “beat” Nvidia; it grows when customers deploy TPUs, custom accelerators, or Ethernet‑based AI fabrics alongside GPUs.
Competitive context
– Nvidia and AMD own the general‑purpose accelerator layer; Broadcom monetizes the fabric, optics, and custom designs customers want in parallel.
– Marvell, Intel, and emerging ASIC houses are credible rivals, but Broadcom’s breadth (switch silicon to packaging) and long relationships with top cloud providers are hard to replicate quickly.
– Foundry and packaging constraints remain critical; favored partners with repeat designs get priority.
Key risks to watch
– Customer concentration: Wins with Google and Anthropic are positives, but dependence on a few large buyers can amplify pricing pressure and road‑map risk.
– Insourcing risk: Hyperscalers periodically weigh bringing more design in‑house; Broadcom must keep demonstrating better cost, speed, and risk management than a DIY approach.
– Execution and supply: Advanced packaging and HBM availability can bottleneck ramps; any slip impacts revenue timing.
– Mix vs. VMware software: Broadcom’s software franchise provides high‑margin stability; a heavy swing toward semis can add cyclicality even as it boosts growth.
What to watch next
– Evidence of production ramps: Packaging allocations, HBM supply alignment, and board‑level designs moving from prototypes to volume.
– Networking share shifts: Announcements or disclosures showing large AI clusters opting for Ethernet fabrics using Tomahawk/Jericho‑AI.
– Unit economics: Signs that custom accelerators deliver better cost per training token or inference for customers—a leading indicator for scale.
– Road‑map cadence: How quickly new tape‑outs move to silicon and whether Broadcom’s AI revenue guide continues to rise.
Bottom line
The Google and Anthropic deals underscore Broadcom’s evolution from a merchant chip vendor into a co‑architect of the AI datacenter. They expand Broadcom’s TAM, reinforce its role in scarce advanced packaging, and attach multiple product lines to every incremental AI rack customers deploy. That combination—volume, visibility, and ecosystem leverage—is why the stock is reacting.
This analysis is for information only and not investment advice.
