Two years ago, Nvidia executed its 10-for-1 stock split after shares touched $1,209 - a moment that made the stock accessible to retail investors and turned the pre-split $1,000 mark into a symbolic benchmark. Today, Nvidia trades near $211 per share, which is the post-split equivalent of roughly $2,110 before the split. The $1,000 line hasn't just been crossed - it's been buried under two years of compounding.
But asking whether Nvidia can reach $1,000 again is the wrong question. The question is whether the company's architecture roadmap puts it on the right side of the biggest transition in AI compute since CUDA itself - and whether the return curve to get there still justifies the allocation it commands in your portfolio today.
The inference inflection point
At GTC 2026 in March, Jensen Huang declared the arrival of what he called the "inference inflection point" - the moment when AI spending shifts from training foundational models to running them at scale for customers. He projected $1 trillion in cumulative revenue from Blackwell and Vera Rubin systems through 2027. That is up from the $500 billion revenue opportunity Nvidia had previously cited for the same two-generation window.
Put plainly: Nvidia is betting that inference will become the dominant GPU workload, and that its full-stack architecture - chips, networking, software - will capture it the way it captured training.
This matters because training and inference have fundamentally different hardware requirements. Training demands massive compute density and interconnect bandwidth - CUDA's native fortress. Inference demands cost efficiency, latency optimization, and the ability to run models cheaper per token. CUDA's moat is formidable in training; it can weaken in inference where the customer's primary decision criterion is price-per-inference, not developer ecosystem lock-in.
Nvidia's response is architectural: Rubin, launching in the second half of 2026, and Rubin Ultra in the second half of 2027, are designed to dominate both sides of the equation. The Vera Rubin CPU-GPU pairing is meant to make inference economics so favorable that switching costs outweigh raw per-token savings on competitor hardware.
Revenue velocity confirms the thesis - for now
The numbers are still compounding at a rate that makes skepticism difficult to sustain on pure growth grounds. Nvidia's Q1 FY27 results, reported in May, showed revenue of $81.6 billion - up 85% year over year, beating the $78.9 billion analyst consensus. Data Center revenue hit $75.2 billion, up 92% from a year ago. Non-GAAP EPS came in at $1.87.
Compare that to Q4 FY26, reported in February, which brought in $68.1 billion - up 20% sequentially, itself a monster quarter. The acceleration from 20% sequential to what Q1 implies is a continuation of Blackwell ramping into full production while Rubin pipelines begin.
Nvidia also announced an $80 billion share repurchase authorization and increased its dividend. For a company that returned $41.1 billion to shareholders in fiscal 2026, this signals management confidence in sustained cash generation - not a sign that growth is peaking.
Demand is not the issue. The question is whether the supply chain and competitive dynamics support this velocity over the next 18 to 24 months.
The dark horse problem in inference
Here is where the frame shifts. Inference is the workload most vulnerable to architectural disruption - precisely because cost-per-token is the decision metric, not ecosystem continuity.
AMD's MI400, launched in 2026, brings 432GB of HBM4 memory to the table. Earlier, the MI350 series already challenged Blackwell on paper with 288GB of HBM3e. AMD doesn't need to match Nvidia on absolute performance; it needs to deliver competitive inference at materially lower cost per unit. That is the dark horse pattern - when the leader is building for the architecture that dominated the last cycle, the challenger wins the next one through cost efficiency.
The inference market won't belong exclusively to Nvidia. Custom silicon from hyperscalers - Google's TPUs, Amazon's Trainium, Microsoft's Maia - already handles a growing share of internal inference workloads. These customers don't need CUDA. They need throughput and control over their own capex.
This doesn't mean Nvidia loses. What it means is that market share dilution in inference is structurally different from the training story. Nvidia doesn't need 92% share to grow its revenue - if TAM expands five times, dropping to 70% still means dramatically higher absolute dollars. The risk is that multiple compression happens faster than revenue growth if the market decides inference economics don't favor Nvidia the way training does.
What the $20 trillion thesis actually means for your allocation
I still believe Nvidia can reach a $20 trillion market cap - but I've said this before, and I'll say it again with the same caveat: much of that 310% return is back-half weighted, likely in 2028-2030, not in the next 12 months.
Nvidia's market cap sits somewhere in the $5 to $6 trillion range today (at roughly $211 per share and approximately 24.9 billion shares outstanding). Getting to $20 trillion requires either sustained triple-digit revenue growth through 2030, or a multiple expansion that assumes inference software monetization succeeds at scale, or both.
The inference software story is real - Nvidia's NIM microservices, AI Enterprise licensing, and the push toward recurring software revenue on top of its installed hardware base are genuine paths to higher multiples. Hardware sets the ceiling; software sets the multiple. But software monetization takes time. It is not yet showing up as a dominant revenue line item, and the market is pricing in success that hasn't materialized in the financials yet.
This is the capital allocation question: if the best returns are back-half weighted, a 10% position starts demanding active management. Trimming into strength, rotating partial proceeds into positions where the return curve is front-loaded, and keeping a core Nvidia allocation for the long arc - that is the framework that matches the evidence.
What would change my mind
I'm not changing my long-term thesis on Nvidia. But here's what would:
- Rubin ships and inference economics don't deliver the per-token advantage Nvidia promises. If customers find that AMD or custom silicon runs the same models at 40% lower cost with acceptable latency, the inference moat argument collapses.
- Data Center growth decelerates below 50% sequentially for two consecutive quarters. The current 92% year-over-year number leaves no room for the kind of slowdown that would break the growth-at-any-price narrative.
- Software revenue doesn't show meaningful acceleration through fiscal 2027. If NIM and AI Enterprise stay below 5% of total revenue by the end of next year, the multiple expansion thesis has no engine.
- Supply commitments surge beyond demand. Nvidia has a track record of matching capacity to orders, but a near-90% sequential jump in supply commitments - the kind we saw earlier in the cycle - is a warning signal even when demand looks robust. It creates leverage risk if the transition slows.
The debate is not about whether Nvidia remains the dominant AI infrastructure company. It is about whether the path to $1,000 per share - a line the stock has already obliterated - matters less than whether the return profile over the next 18 months is still as compelling as what exists elsewhere in the AI trade. I believe the answer depends on which side of the inference transition your portfolio is on.
Nvidia is on the right side. But "on the right side" doesn't mean every allocation size makes sense at every price.