Something doesn’t add up.

AI stocks have pulled back sharply over the past few weeks, caught in a broader risk-off move tied to geopolitical tensions in the Middle East.

But at the same time, the companies powering the AI boom have been reporting some of the strongest numbers – and issuing some of the most aggressive forward guidance – we’ve yet seen in this cycle.

Historically, these kinds of gaps between price action and underlying fundamentals don’t last very long.

Because they can’t both be right.

What will be left when the smoke clears?

A set of AI tailwinds that are still intact, still accelerating – and now trading at a discount after a fear-driven correction.

AI Infrastructure Data Is Telling a Different Story

So let’s talk about those fundamentals. Because I have five major transcripts from the last few weeks sitting in front of me – Broadcom (AVGO), Marvell (MRVL), Oracle (ORCL), Micron (MU), and Nvidia (NVDA) CEO Jensen Huang’s GTC keynote – and they’re all pointing to the same conclusion: the AI infrastructure supercycle is only compounding.

We’ll start with the companies’ forward guidance revisions.

Forecasts Are Moving Higher Fast

Back in September 2025, Marvell told investors that fiscal 2027 revenue would be ~$9.5 billion. By December, it was revised upward to $10 billion. Last week, it hit $11 billion – with fiscal 2028 now targeting $15 billion. That is a 30%-plus upward revision to the forward revenue outlook, all in six months. Marvell’s projected 2027 growth rate is roughly double what it told the Street at the September investor day.

That kind of revision in six months would be a headline in any other environment.

Here, it’s part of a broader pattern across the stack.

Across the AI supply chain, companies aren’t just reporting strong demand – they’re adjusting expectations higher as that demand shows up faster than planned.

Scale Is Expanding Across the AI Infrastructure Stack

Broadcom’s latest results reflect that same shift, just at a different scale. The company reported $8.4 billion in AI semiconductor revenue in a single quarter, up 106% year-over-year, and guided to $10.7 billion next quarter – implying 140% growth.

Then CEO Hock Tan added a longer-term datapoint that’s hard to ignore: Broadcom now has visibility into more than $100 billion in AI chip revenue by 2027. Not total revenue. Just chips.

If Broadcom highlights the scale of what’s building, Oracle offers a view into how far ahead customers are already committing.

Oracle’s remaining performance obligation (RPO) – essentially a signed backlog (contracted demand that still needs to be delivered) – now stands at $553 billion. AI Infrastructure revenue grew 243% year-over-year, while MultiCloud Database revenue grew 531%. 

Supply Constraints Are Already Showing Up

And in some parts of the stack, demand is already running into supply constraints.

Micron reported the largest sequential revenue increase in company history and projected that next quarter’s revenues will exceed the company’s entire annual revenue for every year through fiscal 2024 – with gross margins rising from 75% to 81% in a single quarter.

Those margins reflect how tight supply has become.

Step back, and all of these data points start to line up with what Nvidia is seeing at the system level.

At March’s GTC event in San Jose, Jensen Huang said: a year ago, he saw $500 billion in high-confidence demand through 2026. Today, he sees at least $1 trillion through 2027. And then, just to make sure nobody was getting too comfortable, he added: “We are going to be short.”

Why AI Demand Is Compounding at an Exponential Rate

Individually, those numbers are impressive. Together, they describe a demand curve that’s starting to bend upward.

Jensen Huang explained what’s driving that shift at GTC.

In the last two years, computing demand has increased by approximately 1 million times. That’s the product of two separate multipliers: 

• First, the compute required per inference session increased roughly 10,000x as AI evolved from simpler chatbots into reasoning models (o1, o3) and then into increasingly agentic systems. 
• Second, usage itself has grown roughly 100x. 

Multiply those drivers, and you get a million-fold increase in demand.

The Shift From Training to Inference Is Driving AI Infrastructure Demand

AI no longer just responds. It acts. The critical development Jensen highlighted at GTC is the inference inflection. For the first two years of the generative AI era, most compute demand was training. Now, with reasoning models that think before they respond – and agentic systems like Claude Code that can autonomously read files, write code, test, and iterate – inference is the dominant and rapidly growing workload. 

Every action requires tokens. Every token requires inference, and every inference requires compute, memory, bandwidth, and power. The demand engine has fundamentally shifted from a one-time training cost to a perpetual inference tax on every activity that AI performs.

This is a structural change. And it explains why every company in this stack is not just growing – but growing faster than they were six months ago.

AI Bottlenecks Are Shifting – And So Is the Opportunity

When demand starts compounding like this, something has to give.

In AI infrastructure, that ‘something’ shows up as bottlenecks – and they don’t stay in one place for long.

GPUs and other accelerators were the first constraint, and that part of the market is now well into a phase of sustained hypergrowth.

From Compute to Connectivity

From there, the pressure moved into interconnects – the systems that link all of that compute together.

Marvell’s results make that shift clear. Its interconnect business, which was previously expected to grow in line with overall capital spending, is now growing at more than 50% – much closer to the pace of the accelerators themselves.

Now the bottleneck has moved again.

The AI Infrastructure Bottleneck Has Shifted to Memory

Memory is the current constraint, and Micron’s numbers show just how tight things have become.

The company is only able to meet roughly 50% to 66% of customer demand, as both AI workloads and traditional server demand compete for limited DRAM and NAND supply.

That imbalance isn’t resolving anytime soon.

High-bandwidth memory (HBM4) is only just beginning to ship, the next generation (HBM4E) doesn’t ramp until 2027, and new fabrication capacity takes years to build.

In the meantime, pricing power is doing the adjusting.

Micron’s gross margins jumped from 75% to 81% in a single quarter – an unusually sharp move that reflects how constrained supply is relative to demand. Its CFO Mark Murphy was explicit: this is not a cycle. Memory has been “recast as a defining strategic asset in the AI era.”

Demand Is Getting Locked In Early

As supply tightens, customers aren’t waiting around.

They’re committing earlier – and at larger scale – to secure what they’re going to need.

We can see that shift clearly in Oracle’s numbers. Its $553-billion RPO may be the single most underappreciated number in technology right now. 

Three years ago, Oracle was a legacy database vendor fighting for relevance. Today, it is the preferred infrastructure for large-scale AI training and inference workloads. Nvidia confirmed this at GTC, noting Oracle as its first AI customer and pointing to Cohere, Core, Fireworks, and OpenAI as tenants. Oracle’s bring-your-own-hardware model – $29 billion in new contracts since the last earnings call – allows it to grow without a corresponding free cash flow drag. 

Demand is accelerating. Bottlenecks are shifting. Capacity is getting locked in.

Now the buildout itself is starting to change.

The Rise of Custom Silicon In AI Infrastructure

Both Broadcom and Marvell are seeing the same shift from different angles: hyperscalers are increasingly building their own custom AI chips.

Broadcom is directly exposed to that trend.

The company now serves six XPU customers: Alphabet (GOOGL), Anthropic, Meta (META), ByteDance, Fujitsu, and OpenAI. And importantly, these are multi-year partnerships tied to each company’s long-term AI roadmap.

OpenAI alone has signed a 10-gigawatt agreement through 2029 and plans to deploy more than 1 gigawatt of its first-generation XPU in 2027.

The reason for this shift is straightforward.

As AI models become more specialized – whether for reasoning, inference, or sparse architectures – general-purpose GPUs can’t always deliver the same efficiency as chips designed for a specific workload.

That’s where Broadcom has an advantage. Its decades of experience in custom silicon design, combined with advanced packaging and manufacturing scale, make it one of the few companies capable of delivering these chips at volume.

Marvell sits in a different position – but benefits from the same trend.

Every XPU that gets deployed still needs networking, memory expansion, and high-speed connectivity. Marvell’s portfolio – network interface cards (NICs), CXL-based memory expansion, and switching – supports that layer of the buildout.

As more custom chips are deployed, that “attached” market grows alongside them.

Marvell expects that portion of its business to reach roughly $1 billion by fiscal 2027, with a path to more than $2 billion by 2029 in networking and memory-related products alone.

It doesn’t need to design the winning chip.

It supplies the infrastructure that connects and supports all of them.

Short-Term Noise vs. Long-Term AI Demand

None of the demand trends we’ve just walked through have been driven by geopolitics.

They’ve continued to build in the background.

What the U.S.-Iran conflict has done is introduced a layer of macro uncertainty – pushing energy prices higher, tightening financial conditions, and triggering a broad risk-off move across equities.

The key question is how durable that overhang is.

Right now, the rhetoric remains elevated, and negotiations have been uneven. But the underlying incentives on both sides point in a different direction.

Sustained escalation carries meaningful economic costs – through energy markets, trade flows, and domestic financial conditions – that neither side is well-positioned to absorb for long.

That doesn’t guarantee a clean or immediate resolution. But it does suggest that the current level of geopolitical risk premium is more likely to stabilize or gradually fade as those pressures build.

When that happens, the market’s focus will shift back to underlying fundamentals.

And in this case, those fundamentals have continued to strengthen while attention has been elsewhere.

The stocks that get hit hardest in risk-off moves in a sector with intact fundamentals are typically the same stocks that recover fastest and furthest when the risk-off catalyst resolves. 

What Happens When Price Catches Up to Data

Jensen Huang now sees at least $1 trillion in AI infrastructure demand through 2027 – and expects supply to fall short.

Broadcom is scaling custom silicon programs tied to multi-gigawatt deployments.

Oracle has already locked in hundreds of billions of dollars in future demand.

Micron is operating in one of the tightest supply environments in its history.

The data is already on the table. 

The AI infrastructure buildout is still accelerating. As more compute comes online, the companies positioned on top of it – turning it into products, platforms, and recurring revenue – will begin to take a larger share of the upside.

That layer is coming into view.

And one company, in particular, sits right at the center of it.