Broadcom Isn't Surviving the AI Bubble. It's Riding the Market Shift Nobody's Fully Priced.

Broadcom isn't surviving an AI bubble. It's riding a market structure transition the market hasn't fully priced.

When people call an "AI bubble," they're usually thinking in terms of macro multiples and crowded positioning. UBS sees hyperscaler capex growth slowing to 25% in 2027 and 6% in 2028. Bank of America found 82% of surveyed investors viewed semiconductors as the most crowded trade. The bear case sounds clean. But it's the wrong framework.

The market isn't shrinking — it's bifurcating. The AI compute landscape is splitting from training-dominated to inference-dominated, and the hardware architecture that wins each phase is different. Training still favors Nvidia's general-purpose GPUs with their CUDA ecosystem. Inference — which Deloitte projects will account for two-thirds of all AI compute in 2026 — favors purpose-built silicon optimized for cost, efficiency, and latency at production scale. Custom ASICs are growing 44.6% year-over-year; merchant GPUs, 16.1%. This isn't incremental share erosion. It's a generational shift in where the money flows.

Broadcom is the dominant architect of that shift.

The custom ASIC moat

Broadcom's AI semiconductor revenue hit $10.8 billion in Q2 FY2026, up 143% year-over-year. That followed $8.4 billion in Q1, up 106%. One quarter of AI revenue in FY2026 already surpassed the company's entire FY2024 AI chip run of approximately $12.2 billion. The trajectory is accelerating, not decelerating.

CEO Hock Tan guided full-year FY2026 AI semiconductor revenue to $56 billion — implying roughly 180% growth — and reiterated a FY2027 target of more than $100 billion. The company carries a $73 billion backlog of committed purchase orders, with long-term supply agreements extending into 2028. That kind of visibility isn't hype. It's contracted demand.

Six hyperscaler customers anchor this pipeline: Google (Broadcom's longest-standing partner, with seven generations of co-designed TPUs since 2014), Meta, Anthropic, OpenAI, ByteDance, and Fujitsu. Each chip is engineered to a specific customer's model architecture over an 18-to-24-month cycle. Once that ASIC is locked in and deployed at scale, switching costs are enormous. This is the opposite of the merchant GPU model, where a customer can swap suppliers between generations if pricing or performance shifts. Broadcom's architecture creates structural stickiness that general-purpose vendors can't replicate.

Anthropic is scaling compute from roughly 1 gigawatt in 2026 to more than 3 gigawatts in 2027. OpenAI's first custom XPU is slated for volume deployment in 2027 at over 1 gigawatt capacity. Meta plans "multiple gigawatts" of Broadcom XPUs starting 2027. These aren't pilot programs. They're infrastructure commitments measured in the same unit as small countries' energy consumption.

Broadcom and Marvell together control roughly 95% of the custom AI ASIC co-design market. Counterpoint Research projects Broadcom alone will capture approximately 60% of that custom market by 2027. Custom ASICs are forecast to reach 27.8% of the total AI chip market in 2026, and their 44.6% growth rate is nearly triple the 16.1% projected for merchant GPUs.

The networking multiplier

The custom chip story is only half of Broadcom's AI play. The networking stack is the second — and it compounds.

For every $1 spent on AI accelerators, hyperscalers spend an additional $0.40–$0.60 on networking to connect them. Broadcom dominates AI networking. Tomahawk 6, the industry's first 102.4 Tbps Ethernet switch chip, entered volume production in March 2026. Jericho 4, a 51.2 Tbps fabric chip designed to interconnect over one million XPUs, began shipping in August 2025. Nvidia's competing Spectrum-X1600 isn't expected in volume until the second half of 2026 — meaning Broadcom's networking lead carries through at least the next 12 months.

The networking multiplier means AI chip revenue doesn't exist in isolation. As the $73 billion backlog converts to shipped accelerators, networking revenue follows as a built-in tailwind. This is what separates Broadcom from pure-play chip designers: it earns on both the compute node and the fabric that connects them.

The architecture advantage

Broadcom isn't just faster to market — it's architecturally ahead. The company announced in February 2026 that it began shipping the industry's first 2nm compute SoC, leveraging its 3.5D XDSiP platform. This architecture enables packages exceeding 6,000 mm² of silicon with up to 12 HBM (high-bandwidth memory) stacks, using TSMC's SoIC face-to-face 3D stacking and 2.5D CoWoS integration. Four N2 compute dies, one I/O die, and six HBM modules, all on one platform.

This matters because 2nm isn't just a smaller process node — it's a step change in power efficiency per watt, which is the single most important metric for inference at production scale. When you're deploying compute measured in gigawatts, a 2nm advantage translates directly into lower total cost of ownership. Custom silicon already offers up to a 65% TCO advantage over conventional GPUs for inference. The 2nm leap widens that gap further.

The VMware drag

However, Broadcom isn't a pure-play AI story. The infrastructure software segment — anchored by VMware — grew just 9% year-over-year in Q2 FY2026, down from 13% in Q1 and roughly 25% the prior year. Total infrastructure software revenue was $7.18 billion, missing consensus of $7.32 billion.

This segment carries a 93% gross margin and approximately 79% operating margin. It's the funding engine for Broadcom's capital-intensive AI roadmap. A sustained slowdown in software growth raises a real question: if AI hardware margins (currently around 65%) can't fully replace software margin, does Broadcom need to stretch the balance sheet?

The company has $91.47 billion in total debt against $19.63 billion in cash, leaving net debt of roughly $45.3 billion. Free cash flow is $32.76 billion trailing twelve months — strong, but the debt load gives Broadcom less optionality than a cleaner balance sheet would. Management expects software growth to reaccelerate in Q3 FY2026, with projected sales of $8.9 billion, but the Q1-to-Q2 deceleration pattern has already happened once. I'd want to see two consecutive quarters of software acceleration before declaring the trend reversed.

VMware customers have been migrating to competitors like Nutanix over licensing changes and the forced shift from perpetual licenses to subscriptions. The subscription conversion created short-term revenue bumps — VMware grew 25% to $6.6 billion in FY2025 — but that was migration, not net new customers. Once the conversion tailwind runs out, organic growth is the test.

The supply chain bottleneck

Broadcom flagged supply chain constraints in March 2026, specifically at TSMC. CEO Hock Tan said the company had "fully secured capacity of leading-edge wafers and high-bandwidth memory through 2028," but the bottleneck is real. Nvidia is TSMC's largest customer, and both companies are competing for the same CoWoS packaging capacity and advanced-node wafers.

This isn't a demand problem — it's a capacity allocation problem. When TSMC has to choose between Nvidia and Broadcom for limited 2nm or 3nm capacity, the customer relationship, order timing, and strategic priority matter. Broadcom's $73 billion backlog gives it a strong position, but the constraint means the $100 billion FY2027 target depends on TSMC delivering, not just on Broadcom designing.

Put plainly: the risk isn't demand. The risk is whether the supply chain can execute. And when the supply chain constrains, pricing power goes to whoever has capacity. That's a double-edged sword — Broadcom benefits from scarcity but is also limited by it.

Valuation and the allocation question

Broadcom trades at a market cap of $2.035 trillion, up 40.2% over the trailing 12 months and 23.6% year-to-date. Forward P/S is 27x. Compare that to Nvidia at P/S of 21x, with 70.7% YoY revenue growth and 64% operating margins.

The valuation gap tells a story. Nvidia's revenue growth is still faster, but Broadcom's AI chip revenue specifically is accelerating more dramatically — up 143% year-over-year. The market is pricing Broadcom for the AI chip ramp to sustain and for VMware to recover. It's pricing in the $100 billion AI target.

I believe the long-term thesis is intact — Broadcom is on the right side of the training-to-inference transition, with structural advantages in custom ASIC design, networking, and 2nm packaging architecture. The architecture gap between its XPU platform and Nvidia's merchant GPUs will widen in inference workloads, where cost efficiency matters more than absolute raw performance.

But the near-term return profile is different. The stock has already run roughly 40% this year. At 27x forward sales, much of the 2027 growth path is front-loaded. The VMware drag adds execution risk to a valuation that assumes recovery. TSMC capacity constrains how fast the $100 billion target can materialize.

The debate isn't whether Broadcom remains important in the AI infrastructure stack. It is whether the return profile from here justifies the allocation relative to the alternatives.

If you're looking for the company best positioned on the inference side of the AI compute transition, Broadcom is that company. The architecture, the customer base, the networking multiplier, the 2nm lead — it's the right setup. But I think much of that setup is already digested by the market. The stock price reflects the bull case, not a discounted entry point.

What would change my view on the near term? A material acceleration in VMware software growth would validate the margin thesis and support the current multiple. A sustained TSMC delivery delay would pressure the $100 billion target timeline and compress valuation. A broader hyperscaler capex slowdown — which UBS models at 25% growth in 2027 — would test whether contracted backlog survives budget cuts.

In my opinion, Broadcom deserves a position, but the sizing matters. The architecture is right, but the valuation has front-loaded the payoff. The opportunity cost question — is this capital better deployed elsewhere in the AI stack — is the real one. I believe the long-term AI chip growth path is intact, but the next 30-50% of that return is likely back-half weighted, not front-loaded.