Nvidia Below Sector PE|CoreWeave Lands Anthropic Deal

The Cheap Giant Problem

Nvidia just guided to $78 billion in quarterly revenue, beating Wall Street's estimate by more than $5 billion. Annual net income doubled to $43 billion. Data center revenue grew 75% year over year to $62.31 billion in a single quarter. And yet the stock is down nearly 5% on the year.

That gap between fundamentals and price action is the starting tension here. Because the number that matters most is not the revenue figure. It is the valuation multiple. Nvidia currently trades at a forward P/E of 21.66 times. The semiconductor sector average sits at 23.57 times. Intel, a company losing market share and restructuring its foundry business, trades at 87 times. AMD trades at nearly 30 times. Broadcom at over 24 times.

Nvidia is being priced like a mature cyclical, not a hyper-growth platform. That repricing is not random. It reflects a genuine structural question forming in the market. The question is not whether Nvidia dominates today. It is whether that dominance has a ceiling — and whether that ceiling is closer than the revenue trajectory implies.

Wells Fargo offered a counterpoint. Their analysts suggested that a $1 trillion data center revenue target for Nvidia may ultimately prove conservative. That is a striking claim, but it lands differently when the stock is already trading at a discount to peers who are growing far slower.

The market appears to be pricing in risk that the earnings reports are not yet showing. Understanding what that risk actually is requires pulling back from the headline numbers.

CoreWeave, Supermicro, and Ecosystem Fragility

CoreWeave went public as an AI compute leasing company — essentially a data center operator that rents GPU clusters to enterprises and AI labs. Within weeks of its IPO, it announced a multi-year deal with Anthropic to power AI workloads. The stock jumped 13% on the news.

That deal matters structurally. Anthropic is one of the highest-profile AI developers in the world, backed by billions from Google and Amazon. When it selects CoreWeave as a primary compute provider, it signals that hyperscaler-adjacent infrastructure is becoming a viable alternative to building in-house. CoreWeave's model — aggregate GPUs, lease capacity, absorb the capital intensity — is gaining validation at the top tier of AI development.

But there is a shadow running directly alongside this story. Supermicro, the server assembler that turns Nvidia GPUs into deployable rack-level systems, just had its co-founder arrested. Wally Liaw faces charges of smuggling $2.5 billion worth of Nvidia-powered servers to China between 2024 and 2025. He has pleaded not guilty. The legal outcome remains open.

The financial exposure, however, is not ambiguous. Supermicro derives 71% of its revenue from Nvidia GPU products. It holds no long-term supply contract with Nvidia. That is an extraordinary concentration without structural protection. Analysts at Bernstein described the arrest as raising serious credibility issues. Susquehanna went further, calling for the removal of CEO Charles Liang, who has held the position for 32 years.

The risk that analysts flagged most directly is not the legal case itself. It is Nvidia's response to it. If Nvidia distances itself from Supermicro — even quietly, through allocation decisions rather than public statements — the downstream impact on Supermicro's business could be severe. A company that depends on a single supplier for nearly three-quarters of its revenue, with no contractual guarantee of supply, has limited negotiating leverage in that scenario.

This is where the CoreWeave and Supermicro stories converge. The AI infrastructure buildout is accelerating. But the companies assembling and delivering that infrastructure are carrying concentrated risks that the headline growth numbers obscure.

The Reversal: Who Wins If Nvidia Stays Expensive to Build With

Here is the structural dynamic that tends to get lost in the Nvidia coverage. The consensus framing treats Nvidia as the infrastructure layer — the foundation everything else is built on. That framing is largely correct. But it misses a second-order consequence.

When a single company controls the dominant compute layer, the downstream ecosystem has an incentive to route around it. Not immediately. Not loudly. But consistently, over time, as capital allocation decisions accumulate.

Broadcom and Marvell are the clearest expression of that routing. Broadcom's AI revenues grew 106% year over year. Its Q2 2026 revenue outlook is $10.7 billion. Custom silicon — application-specific chips designed for particular workloads rather than general-purpose GPU computation — is the mechanism. Google's TPUs are built on Broadcom. Meta and others are moving in the same direction. Marvell hit an all-time high as AI optics demand surged, with Barclays raising its price target to $150.

The point most observers are missing is this: custom silicon does not need to replace Nvidia to reshape the margin structure of AI infrastructure spending. It only needs to absorb the incremental growth at the margin. If the next wave of AI workload expansion routes to custom ASICs rather than Nvidia GPUs, Nvidia's revenue can still grow while its share of total AI compute spend quietly contracts.

Micron presents a different but related dynamic. HBM — high-bandwidth memory — is the bandwidth layer that makes GPU clusters actually work at scale. Micron projects $33.5 billion in revenue for Q3 2026, up from $23.86 billion the prior quarter, with an 81% gross margin. HBM supply remains tight. Micron's stock is up nearly 385% over the past year, compared to Nvidia's 55.8%. That divergence reflects a market beginning to price memory as a bottleneck asset rather than a commodity input.

The energy constraint compounds this. Nevada's utility NV Energy has warned it cannot meet its 50% renewable energy target by 2030 without fossil fuels — driven entirely by proposed data center demand. North Carolina is delaying coal plant retirements. PJM, the regional grid operator covering much of the eastern US, is targeting 15 gigawatts of new power generation specifically for the data center buildout. NextEra dropped its 2045 zero-emission goal entirely.

This matters for AI infrastructure stocks because power availability is becoming a hard constraint on data center expansion. Companies that can operate at higher efficiency per watt — or that enable more compute per unit of energy — carry a structural advantage that is not yet fully priced into the sector.

Scenario Branching: Two Paths From Here

There are two credible paths forward from the current configuration, and the evidence does not yet cleanly favor either.

In the first path, Nvidia's valuation discount closes. The forward P/E at 21.66 times is an anomaly for a company growing data center revenue at 75% annually with guidance that consistently exceeds analyst estimates. If geopolitical uncertainty stabilizes — the inflation pressure from elevated oil prices moderates, rate expectations shift — the compression that has been applied to high-growth tech reverses. In that environment, a 46% gap between the current stock price and Wall Street's mean price target of $268.8 becomes the relevant frame. The Wells Fargo thesis — that $1 trillion in data center revenue is a floor, not a ceiling — gains traction.

In the second path, the valuation discount persists or widens. Supermicro's legal exposure escalates, Nvidia distances itself from that relationship, and the credibility damage extends beyond one company to the broader ecosystem of GPU-dependent assemblers. Simultaneously, Broadcom and Marvell continue to absorb incremental AI infrastructure spend through custom silicon, and Micron's HBM pricing power signals that the memory bottleneck is the real constraint — not GPU availability. In this path, the narrative shifts from "who owns AI compute" to "who owns the binding constraints in AI compute." Nvidia remains large. But the margin of dominance narrows in ways that the current earnings figures do not capture.

The energy constraint is the variable that cuts across both paths. If power availability becomes a binding limit on data center expansion within the next two to three years — which the NV Energy warning and the PJM capacity targets suggest is plausible — then the companies best positioned are not necessarily those with the most GPUs. They are the companies building the most efficient workload architectures. That may favor custom silicon over general-purpose compute in ways that take years to show up in revenue figures, but months to show up in capital allocation decisions.

The evidence leans toward continued Nvidia strength at the revenue level, but with a widening divergence between revenue growth and stock performance unless the macro environment shifts. The more durable opportunity, on current evidence, may be in the constraint layers — memory, custom silicon, and power infrastructure — rather than in the GPU platform itself. But that lean depends heavily on whether Broadcom and Marvell can maintain their design win momentum at hyperscalers, and whether Micron's HBM margin profile holds as capacity eventually expands.

Neither path ends cleanly. The AI infrastructure buildout is large enough that multiple winners can coexist. The question is where the incremental dollar of value accretes — and that question is genuinely unresolved.