Subject: Jensen Huang’s two-hour sales pitch disguised as a keynote
The Setup
Picture this: SAP Center, San Jose. March 16, 2026. A packed house of developers, press, and venture capitalists who forgot to breathe between funding rounds. Jensen Huang takes the stage in his signature leather jacket—because God forbid we forget he’s the cool billionaire—and spends two hours telling us the future is already here, it’s just unevenly distributed between people who can afford his chips and those who can’t.
The headline? $1 trillion in orders for Blackwell and Vera Rubin through 2027. That’s not revenue. That’s purchase orders. Jensen wants you to know people are so desperate for his GPUs they’re pre-ordering the next generation before the current one ships. It’s the GPU equivalent of a Supreme drop, except the hypebeasts wear Patagonia vests and have Series B funding.
Last year, Nvidia projected $500 billion. This year? Double it. Why? Because inference. Because agentic AI. Because everyone needs more tokens.
The Token Factory Analogy (Or: How I Learned to Stop Worrying and Love the Hype)
Jensen spent a solid chunk of his keynote talking about “token factories.” Not metaphorically. Literally calling data centers “token factories.” As if AWS and Azure are now Ford assembly lines, churning out tokens-per-second like Model Ts.
Here’s the pitch: AI companies are bottlenecked by compute. If they could just generate more tokens, their revenue would go up. Ergo, buy more Nvidia chips. It’s elegant in its simplicity. It’s also the exact logic a drug dealer uses: “If you could just move more product, you’d make more money. Here, take this pager.”
The kicker? He’s not wrong. OpenAI, Anthropic, Google—they’re all racing to serve more tokens faster because that’s how they bill. Tokens are the new API credits. And Nvidia? They’re selling shovels in the gold rush, except the shovels cost $250,000 each and require a liquid cooling system.
Vera Rubin: The Next Big Thing™
Jensen unveiled Vera Rubin , Nvidia’s next-gen AI system. It’s named after the astronomer who discovered dark matter, which is fitting because most of us will never see one in person.
The specs:
1.3 million components
10x performance-per-watt over Grace Blackwell (the current flagship)
Ships late 2026
Rack-scale system (translation: you need a literal rack to house it)
Jensen claims it’s optimized for “agentic AI”—agents spawning more agents to accomplish tasks. The demo showed AI agents collaborating, delegating, and multiplying like digital rabbits. It was impressive. It was also deeply dystopian if you think about it for more than five seconds.
The subtext: You need our chips to run the future. Not “a future.” The future. Because if you’re not running agentic AI at scale, you’re already behind.
The Groq Acquisition: When You Can’t Beat Them, Buy Them
In December 2025, Nvidia dropped $20 billion to mostly acquire Groq , a startup founded by the people who built Google’s Tensor Processing Units (TPUs). It was Nvidia’s largest deal ever. Yesterday, Jensen showed off the first product: Groq 3 LPU (Language Processing Unit).
The pitch: Groq chips are optimized for low-latency inference, while Nvidia GPUs are high-throughput. Stick them together in a rack, and you get 35x better tokens-per-watt performance. It’s peanut butter and chocolate. It’s also vertical integration disguised as innovation.
Here’s what Jensen didn’t say:
Groq was a competitor
Nvidia just bought the competition and slapped their logo on it
This is the AI equivalent of Microsoft buying GitHub
The Groq 3 LPX rack holds 256 LPUs and sits next to the Vera Rubin rack. It’s meant to “expand memory” and “boost density.” Translation: You now need two racks instead of one. Congratulations, your data center just got more expensive.
Kyber: The Prototype We’ll See in 2027 (Maybe)
Not content with selling Vera Rubin (shipping late 2026), Jensen teased Kyber , the next architecture after Rubin. It integrates 144 GPUs in vertical compute trays to “boost density and lower latency.”
When does it ship? 2027. Via Vera Rubin Ultra.
Let me translate: We haven’t finished Vera Rubin yet, but we’re already pre-announcing the next thing, so you don’t buy from AMD or Intel in the meantime.
This is the hardware equivalent of vaporware. Apple does it with iPhones. Nvidia does it with chips that cost more than a house.
OpenClaw: The Open Source Darling Gets Nvidia-fied
About two hours into the keynote (yes, it was two hours), Jensen pivoted to OpenClaw.
For those who missed the hype cycle: OpenClaw is an open-source AI agent framework created by Austrian developer Peter Steinberger in January 2026. It exploded in popularity because people want self-hosted agents that don’t phone home to OpenAI or Anthropic.
Steinberger joined OpenAI last month (because of course he did), and now OpenClaw is “living in a foundation” that OpenAI “supports.” Which is code for: OpenAI absorbed it.
Jensen called OpenClaw “the most popular open source project in the history of humanity.” Let that sink in. More popular than Linux. More popular than Git. More popular than the fucking HTTP protocol, apparently.
He then unveiled NemoClaw , Nvidia’s “enterprise-ready” version of OpenClaw. It’s a “reference stack” with:
Enterprise security
Privacy protection
Scalable task automation
Translation: We’re packaging OpenClaw with our chips and selling it to corporations.
The pitch: OpenClaw is great, but it’s dangerous. Agents can “communicate externally and execute without intervention,” which is “obviously a problem.” So Nvidia built NemoClaw to make it safe. For a price.
This is classic embrace, extend, extinguish :
Praise the open-source project
Build a “better” proprietary version
Sell it to enterprises who are too scared to use the free one
Jensen even compared it to Windows: “It’s no different than how Windows allowed us to make personal computers.”
Right. And we all remember how Microsoft’s embrace of open standards went.
The Groq of It: What’s Real vs. Sales Pitch
Let’s fact-check the hype:
Real:
Demand for AI inference is exploding
Agentic AI is the next wave (agents spawning agents)
Energy efficiency matters (Vera Rubin’s 10x perf-per-watt is significant)
OpenClaw’s popularity proves people want self-hosted agents
Sales Pitch:
$1 trillion in orders ≠ $1 trillion in revenue
Purchase orders can be canceled. Ask any supply chain manager.“Token factories” are not a neutral term
Jensen is framing compute as a commodity to sell more chips. It’s marketing, not philosophy.NemoClaw is not “making OpenClaw safe”
It’s Nvidia inserting themselves into the stack. OpenClaw works fine without them.Groq isn’t innovation, it’s acquisition
Nvidia bought a competitor and rebranded it. That’s not R&D, that’s monopoly tactics.Kyber in 2027 is vaporware until it ships
Announcing hardware two years out is a competitive move, not a product launch.
The Awkward Shit: Olaf the Robot
Midway through, Jensen demo’d a robot version of Olaf from Frozen. Yes, really.
The robot waddled on stage, had a “conversation” with Jensen, and presumably existed to show off Nvidia’s robotics platform. It was the kind of demo that makes you wonder if someone in the marketing department has a Disney fetish or if they genuinely think a talking snowman will sell chips to enterprises.
Mashable described it as “awkwardly extra interactions.” That’s generous. It was cringe. The kind of cringe that makes you check your watch and wonder if the Uber you called two hours ago is still waiting outside.
The Uber Partnership: Autonomous Rides by 2028
Jensen announced Nvidia partnered with Uber to launch a fleet of autonomous vehicles powered by Nvidia Drive AV software across 28 cities by 2028. Starting with LA and San Francisco in 2027.
Cool. Except:
We’ve heard this before. Waymo, Cruise, Tesla—they’ve all promised autonomous fleets.
2028 is two years away. A lot can happen. Like lawsuits. Or regulators remembering they exist.
Nvidia doesn’t make cars. They make the chips. If the cars don’t ship, Nvidia still gets paid.
Also announced: Nissan, BYD, Geely, Isuzu, and Hyundai are building Level 4 autonomous vehicles on Nvidia’s platform. That’s great. It’s also a press release, not a shipping product.
The Chart Porn: Impenetrable Data Visualizations
Jensen loves charts. Loves them. The keynote was peppered with graphs showing:
Tokens-per-second tiers (free-tier at one end, $150/million at the premium end)
Inference performance benchmarks
Power efficiency curves
Cost-per-token models
It was dense. It was technical. It was also designed to make you feel like you’d be an idiot not to buy Nvidia chips.
Mashable’s takeaway: “Huang peppered a presentation of impenetrable charts with dazzling visions of data centers in orbit.”
Data centers. In orbit. He actually pitched space-based AI compute.
I want to live in the timeline where Jensen’s fever dreams stay in PowerPoint and don’t become billion-dollar initiatives.
Hype, Hardware, and Hypercapitalism
Here’s the uncomfortable truth: Jensen isn’t lying.
Demand for AI chips is real.
Inference is the bottleneck.
Energy efficiency matters.
Agentic AI is the next frontier.
But he’s also running the most successful sales operation in Silicon Valley. Every slide, every demo, every “partnership announcement” is designed to make you think:
“If I don’t buy Nvidia chips, I’m fucked.”
And maybe you are. Maybe inference workloads really do require Vera Rubin. Maybe the future is agents spawning agents in Groq-accelerated racks. Maybe data centers in orbit are the only way to scale.
Or maybe—just maybe—this is a $4.5 trillion company selling the dream of infinite compute to an industry high on its own hype.
Nobody Asked
What happens when the bubble pops?
$1 trillion in orders sounds great until someone defaults. Ask the crypto miners who bought GPUs in 2021.Who’s paying for all this?
Venture capital. Which means: someone else’s money. When the funding dries up, so do the orders.Is NemoClaw open source?
Nvidia says it’s a “reference stack.” That’s not a license. That’s not a repo. That’s marketing.Why is Nvidia buying competitors instead of competing?
Because monopolies are more profitable than markets.What’s the environmental cost?
Jensen mentioned energy efficiency. He didn’t mention total power draw. Vera Rubin’s rack systems will consume megawatts. Where’s that power coming from? (Spoiler: not solar.)
GTC 2026 was a masterclass in corporate theater. Jensen Huang is a brilliant CEO. He knows how to work a crowd. He knows how to frame the narrative. He knows that if you say “trillion dollars” enough times, people stop asking if it’s real.
But here’s the thing: the future isn’t inevitable. It’s not a fixed timeline where Nvidia wins and everyone else loses. It’s a negotiation. And right now, Nvidia is dictating terms.
So when you read the headlines about $1 trillion in orders, remember:
It’s not revenue.
It’s not profit.
It’s a bet.
And bets can go sideways.
First published on Substack.