Broadcom’s AI Chip Forecast Jitters Korea’s Memory Giants

Broadcom’s AI – Broadcom reported surging AI semiconductor demand, but its forecast for fiscal third-quarter AI chip sales came in below Wall Street expectations. The guidance reset expectations fast—hitting South Korea’s Samsung Electronics and SK Hynix—and refocused investo

On June 4, the market didn’t argue with Broadcom’s demand numbers—it reacted to what the company didn’t promise.

Broadcom posted fiscal second-quarter revenue of $22.19 billion. Its AI semiconductor revenue climbed 143% year over year to $10.8 billion, driven by demand for custom AI accelerators and networking chips. But its fiscal third-quarter outlook for AI chip sales—$16 billion—fell short of Wall Street’s $17.2 billion estimate. The shortfall didn’t kill the AI story; it shifted the conversation from raw demand to the economics and timing around getting AI systems built.

That pivot landed quickly. Broadcom fell 12.6% on June 4, wiping $286 billion from its market value.

For investors, the move was less about whether AI chips are being bought and more about what those chips imply for the rest of the supply chain—especially in South Korea, where Samsung Electronics and SK Hynix sit at the center of the memory side of AI infrastructure.

High-bandwidth memory sits behind the promise

AI accelerator systems rely on advanced memory, and the pressure point is high-bandwidth memory. Broadcom’s custom chips—ASICs designed for workloads such as inference—still require that next layer of hardware to make performance possible.

And it isn’t confined to data centers. The source of strain is spreading outward as well. as RTX Spark systems show local AI workloads moving onto developer machines. That matters because every expansion of AI compute—whether centralized or local—creates another set of memory demands that manufacturers have to plan for.

Investors treated Broadcom’s guidance as a signal about expectations, not just demand. In South Korea, the impact played out like a mood swing tied directly to chip sentiment.

The KOSPI fell 5.5% on June 5 after the guidance hit, and Samsung and SK Hynix—together about half of the KOSPI’s market capitalization—turned a global AI-chip reset into a Korea-market story.

The point wasn’t that HBM demand is gone. It was that the market wanted proof—clearer proof—on margins and capacity, and Broadcom’s forecast didn’t provide enough reassurance.

Capacity planning becomes the real battlefield

Samsung and SK Hynix still have to balance where high-bandwidth memory is going. The source describes the competing pull across Nvidia-linked GPU systems, Broadcom-linked custom accelerators, and other AI customers.

If capacity stays tight. the effects can show up in very practical ways: shipment delays. allocation disputes. and pressure to fund new production at the wrong time. Even when long-term demand remains intact. the short-term logistics and the economics of supply can still hurt quarterly results—or trigger investor doubt.

SK Group Chairman Chey Tae-won added another piece to that pressure map. At Computex in Taipei, Chey told reporters that SK Hynix plans to double memory wafer capacity within five years. He also said AI-driven memory tightness is expected to last until at least 2030, as reported by Tom’s Hardware.

That statement effectively keeps the focus on capacity decisions as central to AI infrastructure planning. It also underscores why a forecasting miss from an AI chip supplier can resonate so loudly in memory markets: memory manufacturers don’t just meet demand, they time it.

Why cloud budgets and enterprise tests matter too

Broadcom’s AI chip growth still depends on cloud and technology customers building out data center infrastructure. At the same time, enterprise buyers are testing cheaper models as AI costs rise, including DeepSeek, which the source points to as part of a broader push for cost scrutiny.

If major AI buyers slow spending or squeeze suppliers harder on pricing, the source says Korean memory suppliers could feel pressure before the long-term demand story changes. In other words, the risk isn’t that AI disappears—it’s that budgets tighten sooner than suppliers expect.

That same theme is showing up beyond pure chips in enterprise planning. where hardware availability. cloud services. identity controls. and endpoint management increasingly overlap. The source connects this to AI agent devices for the enterprise. where accelerator availability. memory pricing. and data center timelines can shift as chipmakers adjust forecasts.

No one is saying AI infrastructure demand has cooled. But Broadcom’s selloff showed how quickly questions about capacity, margins, and delivery timelines can travel through a supply chain—and how much that can affect Korea’s memory market before end-demand data changes.

For teams trying to plan AI infrastructure, the timing matters as much as the appetite. If forecasts wobble, hardware availability can change, memory pricing can jump, and data center timelines can get pushed back—long before anyone can say whether the underlying demand curve has actually bent.

Broadcom AI chips custom ASICs high-bandwidth memory HBM Samsung Electronics SK Hynix KOSPI South Korea memory suppliers AI semiconductor revenue fiscal outlook