On a single day in Taipei, Nvidia revealed both ends of the agent era: the cloud machine that runs your AI at industrial scale, and the thin laptop that runs it on your lap with no cloud bill. Here's how fast this is moving.

On Monday in Taipei, Nvidia did something it's never done in one keynote: it revealed both opposite ends of the AI future at once.

At one end is the most powerful computer the company has ever built — a data-center machine engineered for the AI, not for you. At the other is a thin Windows laptop that runs an AI agent right on your lap, all day, with no cloud bill. Same day. Same company. Both shipping this fall.

The throughline is agents — the kind of AI that doesn't just answer your questions but takes a whole multi-step job off your plate and finishes it. This week, that kind of AI got a home at both ends of the pipe: a giant brain in the cloud to run the heavy jobs, and a personal machine to run your own.

Here's how fast — and how affordable — this is suddenly getting.

TLDR: Nvidia launched two chips that bookend the agent era. Vera Rubin is a data-center platform that runs AI agents 10x faster at a tenth of the cost. RTX Spark is a laptop chip that runs agents locally, with no cloud meter. Together they're a snapshot of how quickly handing real work to AI is becoming normal. Here's what each one does — and what to hand over first.

Vera Rubin isn't a gadget you buy. It's a refrigerator-sized rack of chips that lives in a data center — the kind Amazon, Microsoft, and Google run by the thousands. You'll never see one. You just use what it produces, the way you use electricity without owning the power plant. And this week it hit full production.

What it produces is cheap agent work. Nvidia built it to run agents 10x faster at a tenth of the cost per token than last year's hardware. In plain terms: a multi-step job that might have cost a dollar to run now costs about a dime. Cheaper is why you'll actually use it; faster is why it'll actually finish while you're at lunch.

The twist? The hardware got more expensive — a single rack runs about $7.8 million, up from $2.8–6.5 million. But it does so much more per dollar that the work gets ten times cheaper. You pay more for the engine and a fraction for the mileage — and Anthropic, OpenAI, and SpaceX are already first in line.

Then Nvidia flipped to the other extreme. Meet RTX Spark — a Windows laptop chip, co-developed with MediaTek, that pairs an Arm processor with a Blackwell GPU carrying 6,144 cores and a full petaflop of AI performance. That's data-center-class muscle — in something you slide into a bag.

But the headline isn't the specs. It's that it's the first Windows laptop chip that runs Nvidia's full AI stack natively — which means it can run an AI agent on the device itself, all day, with two things the cloud can't offer: no per-task bill, and your data never leaving the laptop. Dell, HP, ASUS, Lenovo, MSI, and Microsoft's own Surface ship machines built on it this fall.

Step back and look at the leap. One Vera Rubin rack delivers around 50 petaflops and costs millions. A single RTX Spark chip delivers about one petaflop — and fits in a notebook. Power that was exotic and data-center-bound two years ago is now both portable and rentable by the task.

That's the real story of the week: the question is no longer whether you can run capable AI. You'll have it both ways — cheap in the cloud for the heavy lifting, bill-free on your device for the always-on stuff. The only thing left to figure out is what to hand it.

Attio is the AI CRM that keeps you ten steps ahead.

Ask Attio anything. Where should I focus? What deals are at risk? Search, update, and create across your customer data.

Ask more from CRM. Ask Attio.

Try Attio for free

The people who pull ahead this year won't be the ones asking "what can I ask AI?" They'll be the ones who already know "what can I hand to AI?"

Don't wait for the perfect machine — figure out which task on your plate is ready to delegate now, and pilot one. We built a prompt that maps it for your exact job.

Try this prompt.

The Agent Handoff Audit →

It asks about your role and your week, then hands back a ranked list of what to delegate first — plus a low-risk way to test-drive the top one this week.

For eighty years we built machines to wait for us. This week Nvidia built two that don't — one the size of a refrigerator, one that fits in your bag. The only question left is what you'll hand them first.

$NVDA ( ▼ 3.62% )

$GOOG ( ▼ 0.76% )