Nvidia reported $81.6 billion in revenue for the quarter ended April 26 - 85% above a year ago, with data center revenue alone doubling to $75.2 billion. That's the number the market latched onto. It's also the number that distracts from what actually changed.
The real story isn't that demand is strong. Jensen Huang has been saying that since 2024. The story is that Nvidia is now fighting on three separate product fronts at once - training GPUs, inference accelerators, and data center CPUs - and has backed all three with $145 billion in supply commitments that lock in its exposure for the next two years.
Demand is parabolic, as Huang put it. But $145 billion in commitments is leverage. And leverage changes the risk profile even when the thesis hasn't.
The front no one expected: CPUs
Nvidia unveiled its Vera CPU at GTC 2026 - an ARM-based server processor with 88 cores that, in early benchmarks, beat Intel's latest Xeon by over 50% and edged AMD's EPYC by roughly 10% in AI-focused workloads. Nvidia projects about $20 billion in Vera and Grace CPU sales, targeting a $200 billion total addressable market.
This is the part the market misses. Vera isn't about Nvidia stealing CPU market share from AMD and Intel. It's about Nvidia completing the full-stack system so that when hyperscalers build their own racks, they buy the CPU too - not just the GPU. That's what separates a component vendor from a platform vendor. A component vendor rents its position; a platform vendor owns it.
Nvidia has been selling GPUs into systems built around someone else's CPU. That arrangement works until it doesn't - until Google's TPUs, Amazon's Trainium, or Microsoft's Maia chips start eating into the GPU slot, and the customer who used to buy the whole rack starts buying only the part that still needs Nvidia. Vera flips that. If the CPU and GPU come from the same architecture, the switching cost for the whole system goes up.
The front Nvidia paid $20 billion to create: inference
Here's where the competitor headline is partially right. Nvidia's second new front is inference - and it didn't build it organically. Last December, Nvidia signed a roughly $20 billion deal with Groq, absorbing the company's low-latency inference chip architecture and engineering talent.
The result is the Groq 3 LPU (Language Processing Unit), which ships in liquid-cooled LPX racks of 256 accelerators. It targets the decode phase of inference - the part where latency and token-generation speed matter more than raw floating-point throughput. Each LPX rack works alongside Vera Rubin NVL72 systems, creating a disaggregated inference architecture where different chips handle prefill and decode separately.
This is the architectural transition that matters. Training is CUDA's fortress - Nvidia still holds roughly 70% of the AI accelerator market, and in training, that share is higher. But inference is where the moat narrows. Inference cares about cost per token, latency, and power efficiency. It doesn't care as much about CUDA. Google and Amazon already run inference workloads on their own custom silicon. If Nvidia had waited for Rubin GPUs alone to dominate inference, it would have been playing catch-up.
The Groq move was a recognition of that. Pay $20 billion now for a specialized decode architecture, or spend the next three years watching hyperscalers run inference on ASICs that don't carry the Nvidia name.
But there's a wrinkle Nvidia itself introduced. The Rubin CPX - a prefill inference accelerator Nvidia announced in September 2025 and touted at GTC 2026 - was quietly pulled from the roadmap shortly after the conference. The solution appears to be pushed to the next-generation Feynman platform in 2028. Groq's LPX took center stage instead.
What this tells me is that Nvidia is choosing specialization over generalization in inference. The Rubin GPU handles training and broad inference. Groq's LPU handles low-latency decode. That's cleaner than trying to make one chip do everything. It also means the inference transition is more complex than "Rubin replaces Blackwell." It's a two-chip system that needs to be sold, deployed, and managed as one.
The front everyone already knows: training GPUs
Rubin itself - 336 billion transistors, 288 GB of HBM4 memory, 22 TB/s of memory bandwidth, 50 PFLOPS of FP4 compute - ships in the second half of 2026. Nvidia has estimated it will sell $1 trillion worth of chips based on Blackwell and Vera Rubin architectures across 2026 and 2027. Q1 fiscal 2027 guidance projects $91 billion in revenue, implying sequential growth from the $81.6 billion just reported.
Training is where the CUDA moat still works. It's where the $1 trillion revenue estimate is anchored. Rubin is a 3.3x throughput improvement over Blackwell, and in a market where hyperscalers are still building out their first major training clusters, that generation gap keeps Nvidia on the procurement shortlist regardless of what custom silicon looks like.
The problem isn't training. The problem is that training revenue growth will inevitably decelerate from these bases. $75.2 billion in a single quarter of data center revenue means you need $300 billion a year just to stay flat. You need $375 billion to grow 25%. The TAM has to keep expanding, and it has to do it while custom silicon eats the inference edge.
The number that changes everything: $145 billion
Now to the supply chain. This is what I check first, because it leads the earnings narrative.
Nvidia's management discussed scaling supply commitments to $145 billion during the Q1 earnings call. That figure represents the total inventory and fabrication capacity the company has pre-paid or pre-contracted to ensure it can meet demand through the Rubin transition. For context, two years ago, these commitments were a fraction of that scale. The sequential increase is of near-order-of-magnitude proportions.
Then on May 27, Huang announced that Nvidia plans to raise its annual spending with Taiwanese suppliers to $150 billion, up from roughly $100 billion. That's the Taiwan semiconductor cluster - TSMC for leading-edge fabrication, SK Hynix and Samsung for HBM memory, and the networking and packaging supply chain that ties it all together.
Here's what these numbers mean: demand is real. You don't commit $145 billion to supply unless you see $145 billion of orders coming back. But you also don't walk away from $145 billion in commitments if demand slows. That's the leverage risk. A near-100% jump in supply commitments from the prior cycle is the kind of move that turns an earnings miss from a disappointment into a margin catastrophe.
$145 billion in supply commitments is a dual signal: it's proof of demand strength and proof of leverage risk. I don't know of a way to separate the two.
There's one more layer. China data center revenue declined approximately 45% year-over-year in the second half of fiscal 2026, falling from an estimated $12 billion. Export controls are biting, and while management frames China as a small fraction of total revenue, a 45% decline in what was once a meaningful growth engine narrows the margin for error on the rest of the numbers.
What this means for allocation
I still believe Nvidia's long-term thesis is intact. The $20 trillion market cap target I've discussed requires the software monetization story - AI Enterprise, DGX Cloud, inference-as-a-service - to scale alongside the hardware. That return curve is back-half weighted, likely concentrated in 2028-2030 when Rubin achieves full ramp and the Vera Rubin platform becomes the reference architecture for AI factories.
But here's the judgment call: Nvidia now sits near $5 trillion in market cap with $145 billion in supply commitments, three product lines that need to execute simultaneously, and a China revenue line that's collapsing. The company is winning. The question is whether the return profile from here - with all three fronts now live and all three requiring flawless execution - justifies the same allocation as when it was just GPUs and CUDA.
The debate isn't whether Nvidia remains dominant in AI accelerators. It's whether the risk/reward from a $5 trillion market cap, with triple the product execution risk and supply commitments that have nearly doubled, is as compelling as what can be found elsewhere in the AI trade. AMD at a fraction of the market cap, gaining inference share. TSMC capturing the supply chain economics. The memory suppliers who don't need three products to succeed - they just need Nvidia to buy.
Demand is not the issue. The issue is whether $145 billion in supply commitments, a three-front product war, and a $5 trillion valuation still justify a large allocation - or whether the return curve has shifted enough to make smaller positions and opportunistic re-entry the smarter play.
I believe much of Nvidia's return is now back-half weighted. That doesn't mean exit. It means size the position to the execution risk, not to the enthusiasm. If Rubin ships on schedule and inference economics hold, the stock still has room. If either slips, the $145 billion in commitments becomes a very expensive lesson.
What would change my thesis? If Rubin achieves its 10x inference cost improvement on real-world workloads, and if Vera CPU adoption among hyperscalers moves faster than expected - that reconfirms the platform thesis and justifies holding through the leverage. If Rubin inference economics fail to materialize, or if supply commitments have to be cut as demand normalizes, that's the inflection point where the thesis requires rethinking.
Time horizon matters. If you're investing for 2030, this is a holding problem, not a selling one. If you're investing for 2027, the risk/reward requires a different allocation entirely.