Hardware Removes Excuses

My $4,000 Nvidia Spark was 2X the speed of my $3,000 M2 Pro - I didn't matter.

It started as a Sunday side project—something to do between the noon Bears game and yet another round of debugging vLLM authentication that night.

I decided to finally turn my Gmail archive into data. Fifteen years of messages. 103,000 emails. Fourteen gigabytes of text that should, in theory, show how I’ve changed: the tone of my writing, the way I negotiate, how curiosity gave way to confidence.

I had the perfect setup. A GB10-based NVIDIA Spark system—an AI-class SoC that can embed a hundred thousand documents in minutes. No token limits. No rate caps. No excuses. By late afternoon, I had everything imported, cleaned, and benchmarked.

The GB10 was roughly twice as fast as my M2 Pro with 16GB. I ran the comparison partly for validation, partly because I’m constitutionally incapable of not benchmarking hardware. The Spark system flew. The MacBook was respectable. Neither speed mattered.

That’s when I ran out of skill.

The stack between the silicon and my analysis worked flawlessly. Ubuntu didn’t crash. Python didn’t complain. The sentence transformers ran clean. The embedding models did exactly what they were designed to do. The infrastructure wasn’t the problem.

The early analytics were easy. Keyword trends, sentiment charts, topic clusters—the usual data-science starter kit. They produced exactly what you’d expect: “AI” spikes in 2018, “cloud” peaks in 2020, sentiment dips during long consulting stretches. Accurate but lifeless. I wasn’t learning anything new—just confirming what I already knew. So I tried something harder.

Conversational posture modeling—a way to measure how my tone evolved. The idea was simple: tag each email as exploratory, authoritative, negotiative, or reflective. See how those ratios changed over time.

In theory, it would chart the arc of my career. In practice, it fell apart instantly. The model tagged an email about a $200K consulting win as neutral/informational while flagging a routine expense report as anxious. It wasn’t wrong—it just had no concept of what mattered.

The context lived in my head, not the text. And no amount of compute could bridge that gap.

That’s when the real insight landed: hardware removes excuses, but it doesn’t create understanding. The GB10 gave me all the speed and capacity in the world. The M2 Pro would have gotten me to the exact same ceiling, just thirty minutes slower. The missing layer wasn’t compute—it was interpretation. I could process the data, but I couldn’t frame it.

I’m not a data scientist, and trying to brute-force meaning out of embeddings only made that clearer.

By the time I closed the terminal—somewhere around 10 p.m., long after the Bears had won and my patience with vLLM authentication had too—I’d embedded a hundred thousand messages on two different systems and learned one honest thing:

The real work starts after the embeddings finish running.