Hock Tan and his CFO Kirsten Spears logged into the June 3 earnings call with numbers that should have satisfied anyone. AI semiconductor revenue hit $10.8 billion in Q2, up 143% year over year, above Broadcom’s own forecast. Full-year AI guidance went to $56 billion. The $100 billion fiscal 2027 target was reaffirmed. By any prior measure, a blowout quarter. The stock dropped 12% in after-hours trading anyway, and by Friday had taken the Nasdaq down with it.
The reason wasn’t the numbers. It was the Q&A.
The first crack in the dyke appeared when Blayne Curtis from Jefferies asked the question every analyst in the room wanted answered but none wanted to ask directly. He’d noticed the 8-K disclosing a long-term agreement with Google and got to the point: “I think there is a lot of concern about share within that customer. I was just kind of curious, now that you have this agreement, maybe you could speak to a little bit more in terms of your confidence.”
[Broadcom Q2 2026 Earnings Call Transcript]
Tan called it “a very, very strong agreement” reflecting “the strength of the partnership we have.” Then he said the quiet part out loud: “We also accept the fact that given the rate of growth of consumption and development of AI compute by our partner Google, that we fully expect that there will be some diversity of sources for them.” The market heard the hedge, not the hype. Broadcom’s biggest customer was spreading its bets across multiple chip suppliers to maintain negotiating leverage, control its own technology roadmap, and avoid the sole-source dependency that hands a supplier pricing power over time. Every analyst on the call started mentally discounting the $100 billion target.
The broadening breach in the dyke came from Ross Seymore at Deutsche Bank, who pressed on margins. Gross margin was falling, and he wanted to know why within semiconductors specifically. CFO Kirsten Spears explained the mix dynamics; ASICs and TPUs carry lower margins, AI networking offsets some pressure. Then Tan stepped in: “Structurally, the semiconductor margins remain very stable and very solid. It is the mix, particularly the mix between software and non-AI to the very, very rapidly growing AI semiconductor, that is just diluting gross margin.”
Translation: the faster Broadcom grows its AI chip business, custom ASICs designed exclusively for individual hyperscaler customers: Google’s TPUs, Anthropic and OpenAI accelerators, Meta’s MTIA chips, the worse its blended margins get. That’s not a temporary problem. Spears had already told investors to model the two segments separately going forward, a CFO’s way of saying the consolidated numbers are going to look worse as AI scales.
Put those two disclosures together and layer on top the Q3 AI guidance of $16 billion, below analyst expectations of $17.2 billion, and you have a selloff. Any one of those three was manageable. Together they told a story: growth may be decelerating at the margin, the biggest customer is hedging, and the business growing fastest is the least profitable. Investors who had priced Broadcom as the pure-play AI infrastructure winner had to reconsider all three assumptions simultaneously, in real time, on a Wednesday afternoon.
What makes the margin confession significant, is Jensen Huang had said precisely the same thing two days earlier at Computex, but said it differently. He described AI factories as the largest infrastructure buildout in human history, with single sites heading toward one gigawatt and capital costs of “$50 billion to $60 billion, and soon it will be $80 billion to $100 billion per gigawatt.” He was explicit about where value accumulates in that world: “If you have one gigawatt of power, then throughput per watt is revenues, because every token is profitable, every token is revenues.”
The unit of competition is no longer the chip. It is the system, watts, racks, networking, cooling, optimized to generate the maximum tokens per dollar of capital deployed. Huang was direct about what Nvidia had become: “A long time ago, Nvidia used to be a GPU company, but over the years we’ve evolved to become a systems company. Nvidia has really started to transform ourselves yet again” into an AI infrastructure company that helps customers build entire AI factories, not just buy servers.
Marvell reported four days before Broadcom and gave the market every reason to separate the two. Its optical interconnect and networking business, the pipes between the engines, held margins firm at 58.9% non-GAAP while revenue grew 28% year over year. CEO Matt Murphy upgraded interconnect revenue growth guidance three times in a single call: “It was beginning of this year 30%, then 50% and now 70%.” He was direct about why: “Our networking products, including interconnect and switching, are driving strong revenue growth as networking becomes increasingly critical with each new generation of AI infrastructure. Now in the early stages of generative AI, the primary focus was on addressing compute and memory bottlenecks. As more complex architectures such as reasoning models and mixtures of experts have begun to deploy, the role of networking has become significantly more important.”
[Marvell Q1 2027 Earnings Call Transcript]
The market didn’t care. When Broadcom fell, Marvell fell with it, down 16.74% on the day, the second largest loser in the semiconductor complex. ARM fell 12.84%. Micron dropped 13.25%. A stock up 145% year to date, with margins holding and guidance rising, got sold in the same undifferentiated sweep that took down every name with “semiconductor” and “AI” in its description. Charu Chanana, chief investment strategist at Saxo, told Reuters the selloff reflected a market that had run too far on a single thesis: South Korea had been “one of the biggest beneficiaries of the AI memory supercycle,” making the entire region vulnerable when investors began to question whether expectations around AI demand had run too far ahead of reality.
The same logic applied in Santa Clara. That indiscriminate selling is the most dangerous signal in the entire episode, not because it was wrong about Broadcom, but because it proves the market is reading the sector as a category rather than a supply chain. It cannot yet separate the engines from the pipes, the compute layer from the interconnect layer, companies whose margins are compressing from those whose margins are not. When wave two, that’s the design services exposure in India and Israel, surfaces in an earnings call the same category logic will apply. Everything will get sold. The distinctions that matter will be the last thing priced.
Tan’s margin compression and Murphy’s margin expansion are the same story told from opposite ends. Broadcom’s compute silicon is diluting blended margins as it scales. Marvell’s interconnect layer is holding margin and accelerating. The chips that move data are worth more than those that process it. It’s where the AI buildout’s economics comes to rest, confirmed simultaneously by two CEOs on consecutive earnings calls. The semiconductor industry has spent two years pricing the AI boom as a chip story. Computex said it was an infrastructure story. The Broadcom and Marvell earnings calls confirmed it with gross margin data in the same week. The selloff was the market updating a model that had assumed the chip was where the money was. Two CEOs in Palo Alto, one in Santa Clara, and Jensen Huang on a stage in Taipei all said it wasn’t. Friday was the market catching up.
Nvidia’s position in this story is more complicated. Nvidia does not compete with Broadcom for hyperscaler custom silicon programs. It sells merchant silicon, H100, Blackwell, Rubin, to anyone who will buy it, amortizing R&D across the entire addressable market and running gross margins above 70% because the same chip ships to thousands of customers simultaneously. The hyperscaler custom silicon buildout at Broadcom exists to reduce Nvidia dependency. Google has been running its own TPUs since 2016 for that reason. Huang’s Computex shift from GPU company to AI factory company is his answer to that displacement risk.
But the direction change reveals something more important about where Nvidia’s actual growth market exists. The AI factory pitch wasn’t aimed at Google or Microsoft. They have their own compute ecosystem. Huang was pitching to the wildcatters: sovereign wealth funds building national AI compute, regional cloud providers, petrostates erecting data centers in the Gulf, CoreWeave, xAI, enterprise buyers who want hyperscaler-grade AI capability without hyperscaler dependency or the engineering organization to build from scratch.
These customers cannot design their own TPUs. They need the whole system pre-integrated and they will pay Nvidia’s margin to get it. The problem is that wildcatter data centers are the least equipped customers to absorb an India shock. A sovereign AI facility in Saudi Arabia or Gujarat doesn’t have Google’s engineering organization to run it. It depends entirely on the Indian execution layer: the cloud operations, DevOps, systems integration, and data engineering talent, that Jane Hsu, founder of Researcher and Research LLC identified as the unmodeled dependency. Nvidia’s fastest-growing customer segment is the most exposed node in the chain nobody is watching.
Whether the contagion spreads beyond Broadcom and Marvell depends on a question nobody on the June 3 earnings call thought to ask. The selloff was wave one, the market repricing the assumption that custom silicon margins would hold as AI scales. Wave two hasn’t hit yet. It exists in the design services layer beneath the chip companies: the Indian and Israeli engineering firms doing the RTL design, functional verification, and physical implementation work that turns a hyperscaler’s AI ambition into manufacturable silicon. Their exposure doesn’t show up in a Bloomberg terminal until a program slips or a hiring freeze surfaces in an earnings call.
Wave three is the hyperscalers themselves. If custom silicon timelines extend because design services capacity is constrained or disrupted, hyperscaler AI infrastructure schedules slip, CapEx efficiency falls, and AI revenue assumptions built into valuations that survived Friday’s selloff come back into question. The containment scenario requires hyperscaler internalization, Amazon’s Trainium, Microsoft’s Maia, Meta’s MTIA, to move faster than the macro risks materialize. The problem is that those internalization programs are using the same Indian and Israeli design services firms. The diversification strategy and the concentration risk share the same human capital base. The cure and the disease have the same address.
Intel makes the geographic concentration argument concrete from both sides simultaneously. Its foundry business is competing, so far without significant hyperscaler traction, to manufacture the custom silicon that Broadcom and Marvell currently dominate. But Intel’s most productive chip design operation outside the United States is not in India. It is in Haifa and Petah Tikva, where the team that designed the Core microarchitecture has worked for decades. Intel’s Israeli engineering centers remain among the company’s most strategically important assets, the kind of concentration that would appear nowhere in a standard supply chain risk assessment because it has never had to.
Israel has been stable enough, for long enough, that nobody modeled what happens if it isn’t. That assumption is now being tested in real time. Kristian Kerr, head of macro strategy at LPL Financial, noted that investors may be underestimating how difficult it could be to restore shipping through the Strait of Hormuz to pre-war levels even if Washington and Tehran reach an agreement. Any initial improvement would come from clearing existing bottlenecks, not a sustained restart in production. The same logic applies to engineering capacity. You don’t restart a semiconductor design program the way you clear a stranded cargo. Two nodes in the custom silicon design chain, India’s implementation layer and Israel’s architecture layer, are simultaneously exposed to the same geopolitical event, and neither exposure appears in a hyperscaler risk filing. Intel knows where its engineers are. Its customers apparently do not think to ask.
What none of the analysts on the Broadcom call mentioned, and the roster was JPMorgan, Jefferies, Deutsche Bank, UBS, Goldman Sachs, Morgan Stanley, Barclays, Bernstein, Melius, Cantor Fitzgerald, Citi, and Charter Equity Research, firms collectively managing trillions in semiconductor exposure, is the layer beneath the infrastructure. Not one asked about the human capital the AI buildout depends on. Jane Hsu, founder of Researcher and Research LLC, draws a line the models miss: the buildout is a physical capacity story, GPUs, data centers, power, networking. Infrastructure, on the other hand is not the same as utilization. It runs on engineering talent, cloud operations, DevOps, data engineering, and systems integration. India is neither an AI consumer market nor a cost-reduction play. It is a structural node in the execution layer the buildout depends on. The risk is not that the data centers fail to be built. It is that implementation, adoption, and enterprise utilization arrive more slowly than the physical buildout assumes.
The chain runs deeper than the earnings calls suggest. The hyperscaler custom ASIC business, the Google TPUs, the Anthropic and OpenAI accelerators, the Meta MTIA chips that Tan enumerated on the call, does not run on American engineers alone. The detailed implementation work between chip architecture and physical silicon flows substantially through Indian design service firms. Tata Consultancy Services, Wipro, HCL, and specialist semiconductor design houses like Sasken and L&T Technology Services do the RTL design, functional verification, physical implementation, and timing closure work that turns a hyperscaler’s AI ambition into manufacturable silicon.
This is advanced semiconductor engineering on leading-edge nodes, and India has built a thirty-year deep capability in precisely these disciplines. When Tata Elxsi and KPIT fell on the NSE in sympathy with Broadcom on Friday, the market was not reacting to sentiment. It was beginning to price the dependency. An India shock doesn’t just slow enterprise IT implementation, it potentially slows the custom silicon programs Tan and Murphy identified as their primary growth engines. The ribbon cuttings are American. The engineering is not.
The data centers get built. The demand they were built to serve arrives late or not at scale. That gap between infrastructure investment and utilization reality is where valuations go to die. Broadcom’s earnings call put a number on it. Hock Tan called it gross margin compression. The market responded with a trillion-dollar selloff. The human capital layer, the engineers in Bengaluru, Hyderabad, Pune, Haifa, and Petah Tikva who turn a data center full of GPUs into a functioning enterprise system, is the part of the story none of them addressed, because none of them has modeled it as a risk.
Nobody rang a bell in July 1997 either.
Also Read:
From the Selfie to Samantha: The Next Trillion-Dollar Behavior
Is Intel About to Take Flight?
Who’s Buying America’s Foundry Future?
Share this post via:
Intel: Pushing EMIB Forward: Design Methodology Insights with Synopsys Tools