Keith Townsend

Joshua 1:6 Be strong and courageous, because you will lead these people to inherit the land I swore to their ancestors to give them.

What we do with our downtime during #CTOARoadtrip

Everyone at GTC was talking about scale. Bigger clusters. Faster GPUs. More memory. Meanwhile, off stage, I recorded a conversation that pointed at a completely different problem. The GPUs aren't saturated. They're waiting on data. We've been optimizing the wrong thing. The constraint isn't compute. It's the data path — how context moves from storage to the model. Fix that, and the economics of every system you're building change. New post breaks down what I saw, why it reminded me of Intel Optane, and what it means for how enterprises should be thinking about AI infrastructure right now. 🔗 Link in below

Intel abandoning Optane may be their worst AI-related fail.

What do you mean today isn't Friday!!! It's been a long week of @Nvidia GTC, and I got a whole lot of opinions.

I've been pricing out a 4-node cluster of @NVIDIA Sparks since the GB300 isn't positioned for non-AI Factory use cases. It's a compromise, but you can get a cluster with 1 TB of RAM for less than $60K. Even a four-node cluster is 512GB, which is impressive. That also brings the Mac Studio M3 Ultra back into perspective. The GB300 pricing brings so much of this into clarity and unfreezes the training lab budget.

GTC wasn't about the enterprise. It's not that the technology wasn't impressive. It's not that NVIDIA doesn't understand AI. It's that none of it was designed for you. The whole event was built around: 👉 Hyperscalers 👉 Neo clouds 👉 Frontier AI builders The companies buying AI factories at scale, priced around token economics. The GB300 tells the story. At ~$170K list, evaluate it as a workstation or an enterprise building block—the pricing doesn't make sense. Limited storage. No scale-out story. Doesn't align with enterprise expectations. But evaluate it through AI factory economics? It makes perfect sense. Priced to token-based value. Avoids cannibalizing larger systems. A contained slice of an AI factory. That's the throughline across all of GTC: → Throughput over integration → Scale over operability → Silence on enterprise guidance None of that is accidental. NVIDIA isn't missing the enterprise. They're sequencing it. First—serve the customers building AI factories. Then—figure out how the enterprise fits underneath. That leaves a real gap. Enterprises are still asking: • How does this fit into what I already have? • How do I operationalize this? • How do I control cost and risk? Those questions didn't get answered at GTC. Because right now, they're not NVIDIA's problem to solve. But they're somebody's problem. New post on The CTO Advisor Substack breaking this down. Link below.

Ran into @bocanuts at the show. I asked him if he had any AMD Strix Halo under that hat.

I don’t typically do these Keynotes but this session with some of the most important companies in AI was too tempting. @cursor_ai @perplexity_ai and the like. #NvidiaGTC

@braysec1 @nvidia When you say NeMo, what do you mean, the model, NemoClaw? Because OpenClaw isn’t 75% of anything production.

@CTOAdvisor @nvidia if Nemo is only 75% as good as Perplexity Computer, it's a no brainer

I called Dell’s SMB line and got pricing for the @nvidia GB300 just shy of $170K. Now put your enterprise cap on. This isn’t general data center spend. This is a project-tied purchase. In most organizations, this goes to VP-level approval. Yes, enterprises will get better pricing. Call it 25%+ off. That doesn’t change the buying motion. This has to be tied to a production outcome. Not a PoC. Not experimentation. And that’s where I struggle with pairing this system with OpenClaw. For where OpenClaw is today, this isn’t the path. If you’re making a $150K+ investment, the value has to come from locality—something you can’t get from a shared pool of GB200-class systems. I’m not seeing that case yet. #NvidiaGTC

I just called @DellTech for a quote on their @NVIDIAAIDev Station. The Dell Pro Max GB300. My feelings are officially hurt. It’s out of the class of system I would buy at this point in my business.

@marcusTh_23 @NVIDIAAIDev @microcenter It is. However the local Microcenter had it for $3,999.

@CTOAdvisor @NVIDIAAIDev @microcenter $4700 is actually competitive for that hardware. Try sourcing components separately.

The number of people dropping $4700 on the spot for a new @NVIDIAAIDev Spark just to walk through the NemoClaw is much greater than Zero. @microcenter has a pretty good thing going. #NvidiaGTC

@AaronErickson @nvidia So after a couple of hours playing with it, I’d have to hold back feedback on NemoClaw itself. OpenClaw is fairly flaky. So, it’s hard to separate the instability of OpenClaw from the governance that NemoClaw promises.

@CTOAdvisor @nvidia Would love feedback!

Bringing my @Nvidia Spark to GTC has a payoff. They have a Claw Bar where I can bring my Spark and do an install and Walkthrough of NemoClaw. #NvidiaGTC

@engineering_bae It wouldn’t hurt at all to interview

I sent a Google Director of Engineering role to a few friends to get insight on it. It was meant to help another friend who is job hunting, but a couple of my friends thought I could do the job and should apply. While I know I wouldn’t be able to land the role (especially in this economy), I realized my mom isn’t the only one who consistently hypes me up for mildly nutso side quests 😂

I’m using WhatsApp via the UI like some animal instead of OpenClaw

AI Factory Economics is simple—if we’re truly building an AI factory, the output will prove it. Faster isn’t better if it’s just false confidence. Show me the economics—real, sustainable results. thectoadvisor.com/blog/2026/01/0…

I honestly believe that the solution is elegant in it specifications. However, it is obvious some guy built it in his basement.

If OpenClaw is truly the ‘most important software since the start of AI,’ and it’s as flaky as it is right now, we’ve got real concerns about enterprise adoption of agentic AI. Hype doesn’t build stable systems.