The $20 Billion CPU Bet Nvidia No One Is Talking About

Nvidia CFO Colette Kress told Yahoo Finance in May that the company expects $20 billion in CPU revenue this fiscal year — combining standalone CPU servers with CPUs integrated into superchips. When she first made that comment, it barely registered in the broader market conversation. The GPU story is louder, and the CUDA moat is the thesis everyone already holds.

But if you actually look at what AI compute architecture is becoming, the CPU renaissance is the single most important architectural shift in the data center since 2022.

The market is preparing for Nvidia's Q2 fiscal 2027 earnings on August 26. Wall Street forecasts roughly $92 billion in revenue, representing nearly 96% year-over-year growth. The stock has rallied 12% over the past two weeks, trading around $217, roughly 8% below its 52-week high of $236.54. The setup looks textbook sell-the-news — the bar is elevated, expectations are loaded, and any stumble in guidance or margin commentary could trigger a sharp reaction.

However, the conversation about what Nvidia is becoming hasn't caught up with the product architecture. Nvidia is no longer just a GPU company. It is positioning itself to own both the CPU and GPU layers of AI infrastructure, and that changes the competitive calculus.

The CPU renaissance is not a revival — it's an architectural transition

Since 2023, the data center story has been simple: GPUs and networking are king. AI training and inference shifted compute demand away from the CPU, and Intel's server business stagnated as a result. That dynamic is reversing, but not because CPUs are suddenly better at AI. The shift comes from the workloads themselves.

Reinforcement learning requires massive CPU clusters. RL environments run on general-purpose processors, and the CPU has become the bottleneck keeping GPUs busy. Agentic workloads, where AI models invoke tools, query databases, and browse the internet, generate far more CPU demand than static inference ever did. And at scale, every GPU cluster needs head nodes — CPUs managing attached accelerators and handling internet traffic while power budgets are prioritized for GPU compute.

The supply side confirms the demand. Intel, which missed the initial AI wave entirely, reported an unexpected uptick in datacenter CPU demand in late 2025 and is now increasing capex for foundry tools while prioritizing server wafers over PC wafers. AMD is projecting server CPU TAM growth in "strong double digits" for 2026. Frontier AI labs are competing with cloud providers for commodity x86 servers. Microsoft's datacenter builds for OpenAI feature dedicated CPU buildings supporting massive GPU clusters — tens of thousands of CPUs keeping the GPUs fed.

This is not a cyclical rebound. It's a structural change in how AI compute is architected.

Nvidia is the only company building for both sides

Here's what separates Nvidia from every other semiconductor company: it's the only one designing CPUs and GPUs as a coherent system.

Grace Blackwell, currently in full production, pairs Nvidia's ARM-based Grace CPU with its Blackwell GPU, using NVLink-C2C to deliver 900 GB/s of coherent bandwidth between processor and accelerator. The CPU memory becomes an extension of the GPU memory. That architecture eliminates the data-movement bottleneck that kills performance in disaggregated systems.

Vera Rubin, the next generation unveiled at CES 2026 and now in full production, goes further. It integrates seven purpose-built chips into a rack-scale AI supercomputer with 220 trillion transistors. The Vera CPU doubles C2C bandwidth to 1.8 TB/s, triples memory capacity to 1.5 TB, and introduces custom Olympus cores with spatial multithreading for 176 threads. It's designed specifically for reinforcement learning orchestration and agentic AI control flow.

Jensen Huang has described the total CPU market opportunity as $200 billion addressable for Nvidia. That's a claim that puts him in direct competition with Intel and AMD — two companies that between them dominate the server CPU market today. But the Grace and Vera architectures aren't trying to win every general-purpose CPU workload. They're optimizing for one thing: keeping Rubin GPUs fully utilized in AI factories.

The numbers behind the architecture shift

The financial proof is in the revenue trajectory. Nvidia's data center revenue has compounded from $35 billion in Q3 2025 to $39 billion in Q4, $44 billion in Q1 2026, and $47 billion in Q2 2026. Each quarter is approximately 10-12% ahead of the last, which is a remarkably smooth acceleration curve for a company generating nearly $200 billion in annual run-rate revenue.

The $20 billion CPU projection is meaningful not just as a standalone number but as a signal of how much infrastructure spending is shifting away from GPU-only designs. In context, Nvidia's total revenue is expected to approach $300 billion for fiscal 2027, making CPUs roughly 7% of the mix. That may sound small, but the CPU layer has historically been an afterthought in AI infrastructure budgets — a commodity cost center. If Nvidia can own that layer alongside the GPU, it means the company captures revenue from both the accelerator and the orchestrator. That's the hardware-to-software value migration in physical form.

The margins tell the rest of the story. Gross margins held at 75% in the most recent reported quarter, operating margins sit at 64%, and free cash flow has compounded 65% year over year to $119 billion trailing twelve months. The balance sheet carries $13.2 billion in cash against $64 billion in debt — with a debt-to-equity ratio of just 4.3%. This is not a company running on leverage. It's a cash machine funding its own next-generation architecture cycle.

What the market might be missing

The consensus revenue forecast of roughly $92 billion for Q2 fiscal 2027 would represent another strong quarter. But the real question isn't whether Nvidia beats on top-line numbers. The question is whether the market is correctly pricing the transition from GPU-centric to system-centric AI infrastructure.

Blackwell is ramping to one million GPUs per month. Vera Rubin is already in full production. The Vera CPU-only servers represent a completely new revenue stream that had zero contribution two years ago. If the CPU market is genuinely $200 billion addressable, as Huang claims, and Nvidia captures even a fraction of that alongside its GPU dominance, the TAM expansion alone justifies the current valuation.

Nvidia trades at roughly 21 times trailing sales, which is rich on a static basis but not unusual for a company compounding revenue at 70% year over year. The PEG ratio sits at 0.30, suggesting that earnings growth is outpacing the multiple. What the market may be underweighting is the competitive implication of owning the full stack. AMD can compete on GPU architecture — its MI400 and MI500 chips are closing the gap on pure compute. But no competitor offers a coherent CPU-GPU-Networking-DPU architecture at the rack scale.

The risk that matters

The sell-the-news risk is real. The stock is 12% higher in two weeks, and Goldman Sachs analyst James Schneider has cautioned that the bar for Nvidia is elevated following the recent move. There's precedent — Cisco dropped sharply despite strong results after a 65% six-month run. If Nvidia reports Q2 revenue in the expected range but guidance for Q3 or full-year fiscal 2027 falls even slightly short of the enthusiasm that's built up, the stock could pull back 10-15% quickly.

The second risk is execution on the CPU side. Grace CPU suffered from branch prediction limitations in its original Neoverse V2 cores, which caused slowdowns in unoptimized HPC and AI workloads. Vera's custom Olympus cores are designed to fix that, but early deployments will reveal whether the architecture actually delivers the throughput needed for RL environments at scale. If the CPU story underperforms, it's a dent but not a death blow — the GPU business is the core engine.

The third risk, and the one I think matters most for allocation, is opportunity cost. Nvidia's stock has returned roughly 22% over the past 120 days and is up about 17% year to date. Those are solid numbers but below what the market has come to expect from this company. The $5.26 trillion market cap already reflects Blackwell's success, Vera's launch, and the CPU opportunity. The remaining upside depends on Vera Rubin adoption accelerating faster than expected, on the CPU market materializing at the scale Huang describes, and on software monetization — which has been a theme for years without showing up in meaningful revenue yet.

Where the capital goes

I believe Nvidia remains on the right side of the AI infrastructure transition. The shift from GPU-only compute to system-level AI factories is real, and Nvidia is the only company positioned to profit from both layers. Vera Rubin represents the first architecture where the data center — not the chip — is the unit of compute, and that will matter more as agentic AI and reinforcement learning scale.

But the question for the next six days isn't about long-term positioning. It's about timing and allocation.

If Q2 revenue lands near the $92 billion forecast and guidance reaffirms the $1 trillion sales run-rate trajectory, the sell-the-news risk dissipates and the stock could retest its 52-week high. If the results are solid but guidance is measured — or if margin commentary reveals any pressure from the CPU ramp — the 12% pre-earnings rally is vulnerable.

In my opinion, the architecture story supports a long-term hold, but the near-term risk/reward ahead of August 26 favors smaller allocations. The $20 billion CPU opportunity is genuine, but it's a 2027-2028 story, not a 2026 quarter story. Trimming 20-30% into the pre-earnings rally and adding back on any post-report dip would align the risk with the timeline of what's actually being priced in.

The debate isn't whether Nvidia stays dominant in AI infrastructure. It's whether the return profile over the next 12 months is still as compelling as what's available in the broader AI trade — where companies trading at lower multiples with equally strong architecture positioning may offer better entry points for the next phase of the cycle.