Oracle vs. Broadcom: The Sorting That Wall Street Saw Coming

Let's talk about what happened when two software titans reported results within hours of each other. On one side, a $16.06 billion quarterly revenue miss by Oracle—down from $16.21 billion consensus, the company's software revenue stumbling to $5.88 billion when analysts wanted $6.06 billion. The stock dropped 13.3% in premarket trading. Sell-offs in Oracle tend to ripple. Salesforce had wobbled on worry that CRM adoption might be cooling. ServiceNow hesitated. The entire ecosystem of enterprise software vendors felt the reverberation, a reminder that cloud adoption was not, in fact, a law of nature.

Then Broadcom reported. Revenue of $18.02 billion, beating $17.49 billion consensus. AI chip sales surged 74% year-over-year to $8.3 billion, crushing internal guidance of $6.2 billion. Management went on the earnings call and, without a trace of irony or caution, told investors that AI chip sales would double year-over-year in the next quarter to $8.2 billion. The stock ticked up 3% in aftermarket trading.

Two data points don't make a thesis. But they do make a mirror.

Oracle was supposed to be the connective tissue of enterprise AI. The database where training data lived. The cloud backend where models would be deployed. For years, the narrative held that any AI boom would benefit—had to benefit—the incumbents. The installed base. The relationship capital. The customer stickiness. Wall Street bought it. Oracle stock made sense in that world.

Broadcom has no enterprise relationships. It doesn't have a moat of sticky contracts or decades of customer lock-in. What it has is something simpler and more durable: the companies building the actual AI infrastructure—Google, Meta, Microsoft, Anthropic, OpenAI—decided they needed custom chips, and they're willing to spend multiyear contracts, tens of billions of dollars, to get them. Broadcom is the manufacturer and designer behind the curtain. The company has a $73 billion backlog for custom chips, switches, and other data center parts for AI over the next 18 months.

That's not a forecast. That's not a TAM (Total Addressable Market) estimate dreamed up by sell-side analysts. That's actual purchase commitments from the hyperscalers.

The shift underneath both earnings reports reveals something historians of technology will eventually recognize as crucial. When AI capex was theoretical—when it was models and blog posts and venture-funded startups with 12-month runways—the entire industry could pretend that software would capture most of the value. Cloud infrastructure would be commoditized. The spreadsheet magic would happen in Salesforce, Oracle, ServiceNow. Enterprise software would democratize AI.

Then reality arrived. Building AGI-adjacent capability requires something like $100 billion in compute infrastructure per company. You can't buy that on AWS and call it a day. You need custom silicon. You need optical switches that talk to that silicon at scale. You need thermal solutions, networking topology, custom ASIC design—all the hardware stuff that enterprise software vendors can't build because it's not what they do. They write middleware. They host databases. They don't design chips.

Broadcom does. In fiscal 2025, AI revenue grew 65% year-over-year to $20 billion, driving semiconductor revenue to a record $37 billion. The company isn't explaining away missed guidance or walking back TAM assumptions. It's raising the dividend 10% and announcing a $73 billion backlog because the customer commitments are so concrete that management has zero uncertainty about whether the AI capex cycle will continue.

This is history sorting itself. The 2024-2025 narrative was that software will eat AI. Oracle, Salesforce, ServiceNow, the whole Magnificent Seven software corps—they would be the beneficiaries of the AI boom. Broadcom was a chipmaker. Unsexy. Cyclical. Subject to the whims of the semiconductor industry's eternal feast-famine rhythm.

What Wall Street got wrong was not the presence of the AI boom. It was the structure of the boom. AI infrastructure isn't being built atop cloud platforms. It's being made by hyperscalers who've decided they can't trust third parties with their competitive advantage. Google made TPUs. Meta built MTIA. Microsoft, OpenAI, and Anthropic all went custom. The only question left is how much of the actual silicon manufacturing and design Broadcom can capture—and the answer, based on the backlog and the quarterly acceleration, is "almost all of it."

The S&P 500 posted another record close. The Dow hit an all-time high. Visa was upgraded by Bank of America on the theory that lower rates mean more borrowing, more spending, more transaction volume. This is not wrong. But it's also not the story. The story is that for the first time in a while, the market is actually being honest about where cash is flowing and why. It's not going to Oracle's database licensing fees. It's going to the people building the actual infrastructure. Broadcom. TSMC. The optical and thermal suppliers that feed the hyperscalers' custom chip production lines.

The narrative collapse of enterprise software in the face of custom AI silicon is the most underrated market event of the year. Oracle's miss didn't just signal one disappointing quarter. It signaled that the entire premise of "AI will be monetized through cloud platforms and enterprise software licenses" was built on sand. The real monetization is happening three layers below, in the spec sheets and backlog lists of companies that make the silicon that makes the models that make the intelligence.

Oracle investors are learning what Broadcom investors already knew: when the customer says they're spending $100 billion on AI infrastructure, you better be sure you're the one building the tools. Not selling the enterprise cloud license. Not renting the database. Building the actual machine.

The market is resorting. Broadcom got sorted into the right pile. Oracle got sorted into the wrong one. Everything else is just color.