556: Thoughts on Nvidia's Roadmap & Competitive Position, Blackwell & Rubin, Intel Consortium?, OpenAI Building the AI Giant, Music, Bad Nobel Prizes, and The Thinking Game

"you couldn’t give Hoppers away"

Growth comes at the point of resistance.

—Josh Waitzkin ♟️

🤐 I had so much (a ton) to say about Nvidia’s recent announcements, their future roadmap, and the implications for the industry’s competitive dynamics that I basically ran out of space in this Edition.

So, I’m cutting this intro short.

Fear not, I’ll be back soon with shower thoughts, recs for random products, and bad dad jokes… 😬

---

🏦 💰 Business & Investing 💳 💴

🏎️ Too Fast, Too Furious: From Hopper to Rubin, Nvidia's 99.97% Cost Reduction Roadmap & More 😎🤖

Above is a super-cut of Nvidia’s GTC 2025 Keynote, condensing 2+ hours into 25 minutes (which is just 12.5 minutes at 2x speed!).

It was a great one, jam-packed with impressive stuff and a few surprises that could turn out to be really important for the future of the industry (we’ll know for sure once they ship). If you’ve got time, watch it—it’s worth it.

Jensen’s “speed of light” approach to product development was on full display.

My main takeaway: Nobody can keep up with Nvidia, much less catch up from behind. As long as they keep executing like this, they’re just too fast. Too furious.

👨‍🍳 The Nvidia Recipe: Jensen Cooks Up 25x Iso-Power Gains While Others Still Prep 🍽️

Let me share my highlights from the presentation and show you the case for this, and let’s see if you agree with my conclusion:

• Blackwell Ultra 300 will be available in the second half of this year. “Blackwell Ultra Tensor Cores deliver 1.5x more AI compute FLOPS compared to Blackwell GPUs or 70x more AI FLOPS for GB300 NVL72 compared to HGX [Hopper] H100.” Memory has been upgraded to “up to 288 GB of HBM3e memory per GPU and up to 40 TB of high-speed GPU and CPU coherent memory per GB300 NVL72 rack”.

Here’s Jensen on how much better Blackwell is than Hopper per unit of electricity:

And remember, this is not iso chips, this is iso power. This is ultimate Moore’s Law. This is what Moore’s Law was always about in the past and now here we are, 25x in one generation as iso power. 

This is not iso chips, it’s not iso transistors — iso power, the ultimate limiter. There’s only so much energy we can get into a data center and so within iso power, Blackwell is 25 times.

This is an important point because, as I often mention, power is the bottleneck 🔌⚡️

This means that even if you could get some other chip that is cheaper per unit of compute, it could still turn out to be more expensive per unit of power.

🔥 40x More Tokens, 50% Less Space, Same Electricity Draw 🔌

The above shows the “token productivity” of the same 100-megawatt datacenter with Hopper vs Blackwell GPUs. Using that same amount of power: 1,400 racks vs 600, yet Blackwell produces 40x more tokens/revenue 🤯

• The next generation after Blackwell will be Rubin and Rubin Ultra in 2026 and 2027, respectively. Rubin will “offer an incredible 50 PFLOPs of dense FP4 compute, more than tripling generation on generation vs B300”. Rubin Ultra will take that to the next level with even larger chips and more memory (3.5x the HBM of regular Rubin), for an incredible “100 PFLOPS of dense FP4” compute 🔥

• The custom Vera ARM CPU will replace the Grace CPU that Nvidia has developed to be paired with its GPUs. Jensen said: “The CPU is new, it's twice the performance of Grace—more memory, more bandwidth, and yet it's just a little tiny 50-watt CPU.” It will have 88 cores, be fabbed on TSMC’s 3nm process, and have fully custom cores rather than ones based on ARM’s Neoverse IP.

• DGX Spark and DGX Station desktop AI computers: They look a bit like Mac Minis, but they pack a punch. The Spark has a GB10 Grace Blackwell and the Station will have a GB300 Grace Blackwell Ultra (delivering 20 PFLOPS and 784 GB of memory).

• Kyber Rack Architecture: Nvidia is making its racks even denser by rotating the cartridges 90 degrees so they are vertical instead of flat. This will help pack even more compute per unit of volume (though power remains the limiting factor almost everywhere). Part of this is possible because of the switch from air cooling to liquid cooling.

• New Spectrum-X and Quantum-X networking chips. They invented a way to have low-power high-bandwidth silicon photonics! 💡

• Nvidia Dynamo open-source software to manage and scale inference datacenters. Jensen calls it “an operating system for AI factories”, and it’s designed to “maximize token revenue generation for deploying reasoning AI models.” It does that by orchestrating inference communication across thousands of GPUs. routing tasks and making sure that utilization is highest while latency is as low as possible. “When running the DeepSeek-R1 model on a large cluster of GB200 NVL72 racks, NVIDIA Dynamo’s intelligent inference optimizations also boost the number of tokens generated by over 30x per GPU.”

This isn’t even everything! ☑️☑️☑️☑️

They also updated tons and tons of software libraries 💾 for all kinds of things — computational biology, self-driving cars, geophysics, robotics, etc — and that’s very important too. It all makes the hardware more useful. They sound very bullish on robotics and announced new ways to train robotics in simulated environments.

This bundle is unmatched. Nvidia customers simply CAN’T get this anywhere else.

Custom chip makers are struggling to keep up with the pace of innovation — by the time they offer something even remotely competitive with one of Nvidia’s chips, the company is already ramping up the next generation and teasing two more after that.

As a buyer of compute, when you see what’s coming down the pipeline, why would you take a gamble on anything else, especially considering that it doesn’t have the software ecosystem and that whatever software exists probably won’t be optimized as aggressively as Nvidia’s?

One of the most remarkable things Jensen said was about the value of Hopper chips (which aren’t that old! All the cutting-edge frontier models are trained on them):

I said before that when Blackwell starts shipping in volume, you couldn’t give Hoppers away and this is what I mean, and this makes sense. If anybody, if you’re still looking to buy a Hopper, don’t be afraid, it’s okay. 

I’m the chief revenue destroyer. My sales guys are going, “Oh no, don’t say that”. 

There are circumstances where Hopper is fine, that’s the best thing I could say about Hopper. There are circumstances where you’re fine, not many, if I have to take a swing, and so that’s kind of my point. 

When the technology is moving this fast and because the workload is so intense and you’re building these things, they’re factories, we really like you to invest in the right versions.

🚤 Nvidia's 900x Speedrun 🚀

If you do the math for Hopper → Blackwell → Rubin:

Blackwell had an up to 68x performance gain over Hopper, resulting in an 87% decline in costs. Rubin is slated to drive even more performance gains – 900x that of Hopper, for a 99.97% reduction in cost.

Isn’t that incredible for a few years’ progress?!

I don’t think he’s really Revenue Destroyer except for his competitors — by improving his products this fast, he’s not only making it hard to compete with, but he’s also keeping the upgrade cycle very much alive.

🧲 It's 1995 All Over Again: Nvidia's Upgrade Cycle Customer Magnet

It reminds me of the 1990s for CPUs — they were getting better so fast that everyone had plenty of reasons to upgrade every few years. Paying up for the latest hot chip was absolutely worth it. The difference in how much faster software ran was not subtle, and every few years you could do things that were previously *impossible* (like 3d games, texture mapping, advanced video codecs, whatever).