Date: 2026-03-17 Event: NVIDIA GTC 2026, SAP Center, San Jose (March 16-19, 2026; keynote March 16) Position: 26.6% (7.12% LEAPS, 19.48% shares) -- largest position Prior thesis: Intact (strengthening) -- sole arms dealer for the AI buildout Stock: ~$183-185 (trading range 180−189 post-keynote)
Holy smokes!! Jensen just told us NVIDIA expects $1 trillion in purchase orders between Blackwell and Vera Rubin through 2027. One. Trillion. Dollars. And then he backed it up with the most impressive product roadmap I have ever seen from any technology company -- seven new chips, five rack architectures, and a clear line of sight to Feynman in 2028. This was not a "next quarter" keynote. This was Jensen planting a flag for the next decade. If you walked into GTC 2026 wondering whether the AI infrastructure buildout is slowing down, you walked out knowing it's accelerating.
Let me lay out what was announced, because the breadth is staggering:
| Product | What It Is | Key Spec | Timeline |
|---|---|---|---|
| Vera Rubin NVL72 | 72 Rubin GPUs + 36 Vera CPUs | 10x inference throughput/watt vs Blackwell, 1/10th token cost | H2 CY2026 samples, early 2027 production |
| Vera CPU Rack | 256 Vera CPUs | 2x efficiency for RL/agentic workloads | 2026-2027 |
| Groq 3 LPX Rack | 256 LPUs (from Groq acquisition) | 128GB on-chip SRAM, 35x inference throughput/MW | Q3 2026 shipping |
| BlueField-4 STX | AI-native storage rack | 5x token throughput, 4x energy efficiency | 2026-2027 |
| Spectrum-6 SPX | First co-packaged optics Ethernet switch | Production now | Shipping |
| Rubin Ultra / Kyber | 144 GPUs in vertical Kyber rack | 100 PFLOPS FP4, 1TB HBM4e per GPU, single NVLink domain | 2027 |
| Feynman | Next-gen platform | New GPU, LPU, Rosa CPU, BF-5 | On track 2028 |
Total platform capacity: 3.6 exaflops compute, 260 TB/s NVLink bandwidth. 100% liquid-cooled, installation reduced from 2 days to 2 hours.
Now here's what really matters from an investment standpoint. Look at the cadence: Blackwell shipping now → Vera Rubin samples H2 2026 → Rubin Ultra/Kyber 2027 → Feynman 2028. That is a new architecture every 12-14 months, each delivering a step-function improvement. The Rubin NVL72 delivers 10x inference throughput per watt at one-tenth the cost per token versus Blackwell. And Rubin Ultra in Kyber doubles the GPU count per NVLink domain from 72 to 144.
I said in June 2024 that "NVIDIA is innovating too fast for anyone to catch within 3 years." GTC 2026 just reset that clock. By the time AMD or custom silicon gets competitive with Blackwell, NVIDIA will be shipping Rubin Ultra. By the time they catch Rubin, Feynman arrives. The innovation treadmill is relentless.
The Groq 3 LPX rack is the first product from NVIDIA's $20B Groq asset purchase in December. 256 LPUs per rack, 128GB on-chip SRAM, 35x higher inference throughput per megawatt. Shipping Q3 2026.
This is strategically brilliant. NVIDIA now has purpose-built inference hardware sitting alongside GPU-based training/inference racks. Customers don't have to choose between NVIDIA and inference-optimized alternatives -- NVIDIA IS the inference-optimized alternative. They closed the one gap in their product line. AWS has already committed to deploying Groq LPUs alongside 1M+ NVIDIA GPUs starting in 2026.
Jensen said infrastructure demand through 2027 will reach $1 trillion, up from his $500B estimate in 2025. Let me put this in context:
Even if NVIDIA captures only 60% of the GPU/accelerator portion (conservative given their dominance), that's a massive demand runway extending well past the "peak 2026" bear narrative.
The token generation math Jensen showed was eye-opening: a 1 GW factory goes from ~2 million tokens/second on Hopper to ~700 million tokens/second on Vera Rubin. That's a 350x increase. And what happens when you make something 350x cheaper? You use a LOT more of it. Jevons paradox in action -- I've been saying this for two years now.
This was the through-line Jensen has been building since the MS TMT conference two weeks ago:
Physical AI requires everything NVIDIA sells -- training compute for simulation, inference compute for real-time decision-making, networking for data movement, and Omniverse for digital twin environments. If this market develops to even 20% of what Jensen envisions, inference demand in 2028-2030 dwarfs current levels. This extends the growth runway years beyond the current training-centric cycle.
NVIDIA released Dynamo 1.0, the "operating system for AI factories," delivering up to 7x inference performance improvement on existing Blackwell hardware without any hardware changes. Adopted by AWS, Azure, Google Cloud, OCI, Perplexity, and Cursor.
This is the CUDA moat deepening. Every layer of software NVIDIA ships makes switching harder. Dynamo manages memory (KVBM), GPU-to-GPU data movement (NIXL), and scaling (Grove). It's the kind of infrastructure software that once adopted, never gets ripped out.
The Nemotron Coalition (Mistral, Cursor, Perplexity, LangChain, and others) developing open frontier models on NVIDIA's stack is another lock-in play. NemoClaw and OpenClaw for personal AI agents running on DGX Station (748GB coherent memory, 20 PFLOPS, supports 1T parameter models locally) further extends the ecosystem from cloud to edge to desktop.
Two partnerships jumped out:
| Metric | Value |
|---|---|
| Stock price | ~$183-185 |
| Market cap | ~$4.5T (post-split adjusted) |
| TTM Revenue | $215.9B |
| TTM FCF | $96.6B |
| Q1 FY27 Guide (ann.) | ~$312B |
| EV/TTM Rev | ~21x |
| P/E (TTM Non-GAAP) | ~38x |
| PEG (73% growth) | ~0.5x |
The stock has been range-bound between 170−190 since late 2024. That's 5+ months of flat price action while the business grew revenue 73% YoY. What we're seeing is textbook multiple compression -- the market is pricing in a deceleration that keeps not happening. At some point, if NVIDIA delivers 70%+ growth through FY27 (and the GTC roadmap says they can), the stock re-rates.
A PEG of 0.5x for the most dominant technology franchise on the planet with a clear product roadmap through 2028 and $1 trillion in demand visibility... that is not expensive. The stock is trading at roughly 15x forward earnings if they grow 70% this year. For context, the S&P 500 trades at ~21x.
The Groq integration. Closing the inference-only gap before competitors could exploit it. AWS deploying 1M+ NVIDIA GPUs plus Groq LPUs is the kind of platform win that compounds.
The roadmap density. Seven chips, five racks, and a clear line to Feynman 2028. No one else on Earth has this breadth. Not AMD, not Intel, not Google's TPU team, not Amazon's Trainium team.
The $1T demand signal. Jensen doesn't throw numbers around carelessly. He's been remarkably accurate on demand forecasts since the AI cycle began. When he says $1T, I take it seriously.
Physical AI substance. This isn't a slide deck. Uber robotaxis in 2027 with named cities and timelines. 18M vehicles/year from OEM partners. Disney building robots on the platform. The demand driver beyond training is materializing.
Vera Rubin execution. Samples H2 2026, production early 2027. Any delay = stock pressure. But NVIDIA has delivered every architectural transition on time since Blackwell. Track record is A+.
Groq 3 LPX shipping Q3 2026. First real product from the acquisition. Needs to deliver on the 35x inference throughput/MW claim. If it does, it's a game-changer for inference economics.
Gross margin through the Rubin transition. Each architectural transition brings margin uncertainty (remember Q1 FY26 at 61% GM during Blackwell transition). Management committed to "mid-70s" long-term. HBM4 pricing from SK Hynix and Samsung is the variable.
Hyperscaler CapEx durability. $1T demand assumes hyperscalers keep spending. AWS committing to 1M+ GPUs is a strong signal, but watch Meta, Google, and Microsoft capex guidance through 2026.
Hold. No change to allocation. Position remains 26.6%.
GTC 2026 was the most comprehensive product launch event I have ever seen from NVIDIA -- and I've been watching these since 2018. The combination of Vera Rubin's 10x cost-per-token reduction, Groq LPU integration, Kyber rack architecture, and the physical AI expansion makes the "peak 2026" narrative look increasingly detached from reality.
I'm not adding here because the position is already enormous at 26.6% and the stock isn't at a fat-pitch level for LEAPS. If we see a 15-20% pullback to $150-155 range on any macro rotation, that would be a screaming LEAPS opportunity. But absent that dip, this is a hold-and-let-it-compound position. The business is growing at 70%+ with 50% FCF margins and a clear product roadmap through 2028. There is no other company like this.
Thesis status: Intact -- strengthening further. GTC 2026 extended the demand visibility, widened the competitive moat, and added the Groq inference dimension. The sole arms dealer just added more weapons to the arsenal.
We will see what happens...
GR
Sources: NVIDIA GTC 2026 keynote (March 16, 2026), TheNeuron.ai GTC summary, Tom's Hardware live blog, DataCenter Knowledge, CNBC, ServeTheHome live coverage, NVIDIA Blog, prior GauchoRico analyses (Q4 FY26 earnings review, MS TMT conference review).