Broadcom 16B AI Guidance|Stock Down 21% After Jalapeño Chip Debut

Chapter 1: The Earnings Paradox That Jalapeño Was Built to Resolve

Broadcom reported AI semiconductor revenue of $10.8 billion in its most recent quarter, up 143% year over year, and guided the current quarter to $16 billion — a figure that implies over 200% year-over-year growth.

The stock fell 12.6% on June 4, the day after that report, and has continued declining to sit roughly 21% below its June 2 record.

That combination — revenue guidance accelerating into triple-digit growth while the stock surrenders a fifth of its value — is not the typical post-earnings drift of a company missing numbers.

The consensus explanation is valuation compression: Broadcom had run 55% in two months ahead of earnings, pushing its price-to-earnings ratio to 81, and investors took profits when the AI revenue guide came in at $16 billion rather than the $17 billion analysts expected.

That explanation is arithmetically accurate but analytically incomplete.

Embedded in the selloff is a second signal: news that long-term custom silicon customers Alphabet and ByteDance are separately in talks with MediaTek and Qualcomm to help design future AI chips, suggesting a desire to reduce single-supplier reliance.

The market read that as a structural threat to Broadcom's custom silicon monopoly — the very business that generates most of its AI revenue.

Today's Jalapeño announcement is Broadcom's answer to that read.

What Jalapeño does to the thesis depends on what it actually is — and that distinction determines whether Wednesday's 1.6% gain is the beginning of a recovery or a dead-cat bounce in a compression cycle.

Chapter 2: What Jalapeño Is and Why the Market Is Missing the Cost Argument

Jalapeño is an application-specific integrated circuit, not a graphics processing unit.

The distinction matters because the two chip types have fundamentally different cost structures at inference scale.

A GPU is a general-purpose parallel processor. When running AI inference — generating a response to a user query — a GPU typically achieves 60 to 70% utilization, because inference is constrained by how fast model weights can move between memory and compute cores, not by raw calculation speed. The remaining 30 to 40% of the chip's compute capacity sits idle, burning power for nothing.

Jalapeño was designed from scratch to eliminate that idle fraction. OpenAI hardware lead Richard Ho stated that the chip was engineered around the specific memory movement, networking, and serving patterns that frontier AI models actually use, targeting utilization close to the chip's theoretical limit.

Broadcom CEO Hock Tan told Reuters the chip delivers performance on par with Nvidia's Blackwell processors. He separately told Bloomberg the figure amounts to roughly 50% cost savings per inference token versus current GPUs.

OpenAI's own announcement was more hedged: it described performance per watt as "substantially better than current state-of-the-art" while noting benchmarks are still being measured and a full technical report will follow.

That gap — a 50% cost reduction from the manufacturer's CEO versus a carefully worded claim from OpenAI itself — is the buried assumption the market has not resolved.

If the 50% figure holds in production, Broadcom's inference ASIC business is not a diversification hedge against Nvidia; it is a separate, faster-growing revenue stream serving a different cost point.

Broadcom's AI revenue at $10.8 billion last quarter came entirely from custom ASICs for hyperscalers. Jalapeño adds OpenAI's inference cluster to that client list, with Hock Tan already exceeding his prior forecast of deploying 1.3 gigawatts of chips next year.

The market's post-earnings reading treated Alphabet and ByteDance's diversification signals as customer-loss risk. That reading assumed Broadcom's hyperscaler relationships are stable partnerships that can fragment. The Jalapeño deal suggests the opposite dynamic: OpenAI, the largest AI compute buyer in the market, went to Broadcom rather than building alone.

Chapter 3: Marvell's 218% Run vs. Broadcom's 10% — Who Is Positioning for the Right Thesis

The investor positioning split is more revealing than the stock price chart.

Marvell Technology has risen 218% year to date. Broadcom is up 10%. Both companies design custom AI accelerators and data-center silicon. Both are riding what has been called the custom silicon wave in AI infrastructure.

The gap is not explained by fundamentals alone.

Broadcom's AI revenue of $10.8 billion last quarter dwarfs Marvell's entire quarterly revenue base of $2.4 billion. Broadcom guided $16 billion in AI revenue for the current quarter; Marvell guided $2.7 billion in total revenue.

The explanation is composition. Broadcom's AI semiconductor business is embedded inside a larger company that also generates revenue from networking silicon, broadband, storage, and infrastructure software including VMware. A dollar of AI revenue growth moves Broadcom's consolidated story less in percentage terms. That dilution is a feature for a diversified investor and a friction for a momentum investor.

Marvell, with a pure-play data-center profile and a beta of 2.28, amplified the same AI tailwind into a larger percentage return. Retail investors followed that leverage. A widely shared Reddit thread captured the sentiment: a poster who bought Marvell at $82 watched it reach $325, unable to decide whether to sell.

The problem with that framing is valuation. Marvell now trades at a trailing price-to-earnings ratio of 105 times. Broadcom trades at 65 times — elevated, but 40 points lower on a company with a larger absolute AI revenue base.

If the inference era expands the total addressable market for custom ASICs, Broadcom's diversification is not a drag — it is the reason the company can absorb a 21% correction and still have an AI revenue line growing at over 200%.

The market positioning in Marvell reflects the assumption that pure exposure to the same wave is worth a 40-point premium. That assumption requires that Marvell sustains its current growth trajectory without the kind of post-earnings reset that hit Broadcom when AI guidance missed by one billion dollars.

Broadcom's post-Jalapeño recovery thesis requires a different assumption: that the customer-diversification fear that drove the post-earnings selloff was wrong, and that signing OpenAI as a custom silicon client on a multi-generation roadmap is worth more than one missed quarterly estimate.

Chapter 4: The Verification Window and the Decision Each Investor Class Faces

One credible counter to the Broadcom recovery thesis came from OpenAI itself: the company was careful not to endorse Tan's 50% cost savings figure directly, and a full technical report on Jalapeño's benchmarks has not yet been released.

OpenAI President Greg Brockman affirmed that Jalapeño is not a Nvidia replacement but an addition to existing infrastructure — framing that is honest about near-term limitations but also removes the narrative of a clean Nvidia displacement story.

That is not a fatal flaw in the thesis. The AVGO bull case does not require Jalapeño to displace Nvidia. It requires Jalapeño to add a high-margin inference revenue stream that scales from initial deployment late this year through a ramp in 2027 and full production in 2028.

Tan's stated timeline is concrete: small prototype deployment in late 2026, meaningful scale in 2027, full production in early 2028.

The verification event is specific. When Microsoft and other data center partners begin integrating Jalapeño-powered racks by end of 2026, deployment velocity and performance data will either confirm or contradict the 50% cost savings claim.

If confirmed at scale, Broadcom's forward price-to-earnings ratio of 32 — down from 37 before the selloff — prices in the existing hyperscaler business without attributing meaningful value to the OpenAI inference ramp.

The holder's question is whether to add into the current 21% correction before deployment data arrives. The risk is that the benchmark gap between Tan's 50% claim and OpenAI's hedged language widens in the technical report — the stock would reprice lower on a narrower cost advantage.

The watch-list candidate's question is different: whether the inference era expands Broadcom's addressable market faster than customer-diversification risk erodes it. That answer arrives on the same timeline as the deployment data.

Both classes should watch one variable before acting: the first Jalapeño performance benchmark in production, expected alongside or shortly after initial Microsoft data center deployment in late 2026.

The setup is not a buy signal or a sell signal. It is a hold-with-a-trigger: the stock's recovery thesis is grounded in a concrete claim — 50% cost savings — that has a concrete verification date and a concrete location where it will either be confirmed or contradicted.