From Broadcom to Marvell: How Google's chip internalization is rewriting the ASIC duopoly competition landscape?

On June 3, 2026, after Broadcom (NASDAQ: AVGO) released its Q2 earnings report in which nearly all figures beat expectations, the market’s reaction was unusually intense—after-hours shares plunged 12%, the next day saw an intraday low with a maximum decline of 15%, the stock closed down 12.59%, and the company’s market value evaporated by more than $32 billion in a single day. However, this was not a flawed earnings report: total revenue for the quarter was $22.19 billion, up 48% year over year and above analysts’ expectations of $21.3 billion; adjusted earnings per share were $2.44, beating the $2.40 consensus estimate; and AI semiconductor revenue surged by 143% year over year to reach a record high of $10.8 billion.

This is a textbook “expectation gap” trap. When a stock’s valuation already embeds “all possible optimistic expectations,” then any information that fails to top expectations—even if it matches them—can trigger a repricing of the valuation. And the pressure facing Broadcom does not stop there. Almost at the same time, Macquarie analysts downgraded Broadcom from “Outperform” to “Neutral,” cutting the target price from $513 to $437. The core reason pointed directly to Google accelerating its internalization of chip design and its strategy to diversify suppliers.

The Back Side of a “Perfect Earnings Report”: Dual Squeezing by Expectation Gaps and Valuation Premiums

The numbers in Broadcom’s Q2 earnings report themselves were almost flawless. Semiconductor solutions revenue was $15.01 billion, up 79% year over year, surpassing analysts’ expectations of $14.65 billion; non-GAAP net income was $12.07 billion, up 55% year over year; and free cash flow was $10.26 billion, representing 46% of revenue. Infrastructure software business grew 9% year over year to $7.18 billion. During the earnings call, CEO Hock Tan described AI demand as “simply insatiable,” confirmed that Broadcom is working with six hyperscale customers including Anthropic, Google, Meta, and OpenAI, and announced collaborations with Apollo and Blackstone to build a 20 million kW AI computing platform.

However, market attention centered on three metrics that did not exceed expectations.

Metric 1: Q2 AI revenue below buyer expectations. Broadcom’s Q2 AI semiconductor revenue was $10.8 billion—while it exceeded the company’s own earlier guidance of $10.7 billion, it came in below buyers’ expectations of $11.3 billion.

Metric 2: Q3 AI guidance clearly below buyer modeling. Broadcom expects Q3 AI chip revenue to be $16 billion, growing by more than 200% year over year, but that figure is about $1.2 billion lower than analysts’ models of $17.2 billion. Remember, Broadcom’s stock had already gained more than 40% earlier in the quarter; the valuation multiple’s implied assumption was that “every guidance must break the ceiling,” yet Broadcom’s guidance still failed—on a quantitative basis—to reach the market’s most aggressive expectations.

Metric 3: Long-term AI targets kept unchanged. Hock Tan reiterated the long-term target of “over $100 billion” in AI semiconductor revenue for fiscal year 2027. There was no raise; the figure is also above the market consensus of $57.6 billion for full-year 2026 AI revenue, yet it trails the $125 billion forecast from institutions such as Goldman Sachs. What especially worries the market is that Broadcom did not follow peers like Marvell Technology’s approach of issuing more optimistic long-term guidance; this conservative posture, in Wall Street’s view, intensifies concerns about insufficient long-term growth momentum.

For a stock that had already surged significantly before the earnings release and is trading at historical valuation highs, simply maintaining expectations is not enough—the market wants “to beat expectations.” Wedbush summarized this insight succinctly in its analysis: “Valued at 25–30x forward revenue, the market does not need confirmation—it needs acceleration.”

The macro environment also cannot be ignored. One day before the earnings report, the U.S. Department of Labor released April CPI data. Overall CPI rose 3.8% year over year, above the 3.7% market expectation and also above March’s 3.3%; core CPI rose 2.8% year over year, likewise exceeding the 2.7% forecast. Sticky inflation combined with higher energy costs driven by the Israel–Iran geopolitical conflict further reinforced expectations that the Federal Reserve will keep its “higher for longer” interest rate stance. In a high-interest-rate environment, high-valuation growth stocks typically come under pressure first; and Broadcom’s trailing price-to-earnings multiple of roughly 87 (as of before the report) makes it one of the most sensitive targets.

Google Chip Internalization: The Structural Turning Point of the ASIC Duopoly

If the disappointment at the Q3 guidance level and the unchanged 2027 target are failures from a short-term catalyst perspective, then the chip internalization strategy from Broadcom’s largest customer, Google, creates structural pressure on Broadcom’s ASIC business.

Google and Broadcom’s cooperation on TPUs has lasted more than a decade. However, the past single-procurement arrangement is being replaced by a systematic diversification strategy. Today, Google’s custom chip supply chain has taken on a three-legged structure: Broadcom is responsible for high-performance TPU variants; MediaTek provides cost-optimized “e” series TPUs (20%–30% cheaper); and TSMC is responsible for manufacturing. In recent years, MediaTek has been actively moving into the AI ASIC market. Although it has not officially disclosed whether it won Google orders, industry consensus generally holds that MediaTek has made substantial progress in the race for TPU orders.

What hits the market even more directly is that reports show Google has commissioned Marvell Technology to design customized network chips for its next-generation TPUs. The chip will use Intel’s advanced 18A or 18AP process technology, with plans to enter mass production by the end of 2027, and will be paired with the Humufish TPU designed by MediaTek. Network chips play a connecting role within AI data center clusters: they coordinate data flow between chips, handle congestion, synchronization, and latency—strategically, their importance is no less than that of the main compute chips themselves. This is not the first time Marvell has worked with Google on AI chips; in addition, since 2024 there have been ongoing rumors that Google is actively seeking to collaborate with Marvell to develop AI inference chips, systematically driving supply chain diversification.

Macquarie analyst Arthur Lai provided clear quantitative quantification of this in his rating report. The firm predicts that Broadcom’s share of revenue tied to Google’s TPU will fall from about 95% in 2026 to 80% in 2027, and further to 65% in 2028. Based on that proportion, Broadcom’s ASIC market share loss attributable to just Google as a single customer could amount to at least the hundreds-of-millions to billion-level. Macquarie also cut its 2028 earnings per share forecast for Broadcom by 21% and warned that intensifying competition in the AI ASIC market could, in the long run, suppress Broadcom’s growth and profitability.

From Google’s perspective, the motivations to push chip internalization and diversification are not hard to understand. Broadcom charges a licensing fee per TPU unit. As the scale of TPU demand grows exponentially, the amount Google pays to Broadcom will also surge sharply. Building in-house chip capabilities and bringing in more external partners is the inevitable path for Google to gain control over the AI computing power supply chain’s decision-making voice.

At the same time, Marvell Technology is increasingly becoming another pole in the ASIC customization competition. Nvidia CEO Jensen Huang publicly said at the Taipei International Computer Show that Marvell Technology is “the next trillion-dollar company,” directly boosting market confidence in the broader custom AI chip market. Marvell’s custom data center chip business has become its fastest-growing segment, and its recent Q2 revenue guidance midpoint of $2.7 billion indicates that its expansion pace in the ASIC track has clearly accelerated.

It is important to note that Google’s “de-Broadcomization” is not a direct replacement, but a diversification of the supply chain. Broadcom remains the design partner for the key high-performance TPU variants, and its cooperation agreement with Google has been extended through 2031, so the revenue base in the near to medium term remains relatively stable. But the structural trend has already been established: leading hyperscale cloud providers are shifting from relying on external ASIC suppliers’ external designs to a multi-supplier model with internalized design and diversified procurement—and the impact on Broadcom’s long-term ASIC business moat cannot be underestimated.

AVGO Plungles 15%: A Data Panorama Amid Differences

After the earnings release, Wall Street’s views on Broadcom diverged sharply, reflecting the market’s tradeoff judgment between short-term AI demand and long-term market share.

Sell-side divergence: Bullish camp amid downgrades.

Macquarie’s downgrade to “Neutral” and the target cut from $513 to $437 are the most prominent bearish signals in the current market. Meanwhile, investment banks such as Morgan Stanley, Citi, and Wells Fargo maintained Buy ratings, and Jefferies even raised its target price to $550. Currently, the overall analyst rating consensus for Broadcom is “Moderate Buy,” with an average 12-month target price of roughly $473 to $486—still offering about 11%–16% upside from the $418.91 close after the plunge.

Buy-side logic: Long-term bullishness is not based on short-term guidance.

The bullish reasoning from multiple bullish institutions is not built on an optimistic interpretation of the Q3 guidance, but rather on two core premises. First, Goldman Sachs revised its AI revenue forecasts for Broadcom for 2026 through 2028 to $57 billion, $133 billion, and $193 billion respectively, believing Broadcom’s management could cover cumulative AI revenue of more than $40 billion before 2030. Second, as the current major customers’ XPUs ramp fully in 2027 to 2028, Broadcom’s AI revenue growth inflection point will mainly occur after 2027. After the plunge, Morningstar analysts also sharply raised their fair value assessment for Broadcom from $550 per share to $650, arguing that the valuation based on 18x consensus earnings in 2028 is attractive.

Insider trading signals.

Worth noting is that over the past three months, Broadcom insiders sold approximately $356.4 million worth of stock. Although insider selling can have a variety of reasons and may not directly map to changes in company fundamentals, in the sensitive moment of a post-earnings plunge, this data inevitably prompts the market to re-examine management’s confidence.

Marvell’s two-sided situation.

In this wave of shock, Marvell Technology has also experienced significant volatility. On earnings day, Marvell’s stock fell by about 7%, but the decline was smaller than Broadcom’s. Then, on the day after the plunge, it saw a partial rebound due to news about Google network chip orders. From a structural standpoint, Google’s supplier diversification is clearly a net positive for Marvell: as one of the other poles in the ASIC duopoly, Marvell’s share in Google’s chip design orders is rising. The problem is that Marvell’s Q2 financial report shows its quarterly revenue is only around the $2 billion range, while Broadcom’s AI business alone already reached $10.8 billion in a single quarter. Even if Marvell secures incremental orders from Google, the magnitude gap versus Broadcom in AI revenue scale remains huge. Marvell’s potential upside benefit is still mainly structural supplementation, and it is unlikely to form a true competitive substitute in the short term.

Conclusion

The shockwave triggered by Google’s chip internalization is not the only issue Broadcom faces in the short term. But when the time window is extended, the core driving force behind the current plunge is the “expectation gap,” not “fundamental deterioration.” Broadcom’s AI business will only truly enter large-scale volume ramp after 2027; by then, deployments of XPUs from multiple hyperscale customers will be fully underway. For now, the market is largely pre-digesting the disappointment caused by “no upward revision to the 2027 guidance.”

Macquarie’s target price of $437 reflects a pessimistic pricing of long-term market share loss, while the bullish camp’s target range of $485–$550 reflects sharply different long-term expectations. The crux of the disagreement between both sides is this: at what pace will Google’s supply chain diversification and internalization progress—through gradual dispersion or accelerated substitution?

For investors focused on AI semiconductors, this sharp selloff provides a lens on the core issue: it is no longer whether the Q3 guidance is strong enough, but how deep the long-term moat of the custom ASIC business truly is amid the internalization wave by hyperscale customers.