A little over a month ago, I resigned from my job as a Senior SDE at AWS. Resigning from a high-status, lucrative role that I genuinely enjoyed—and was performing well in—might sound a bit crazy, so let me attempt to justify myself in this post.
First of all, sorry to disappoint: this isn't going to be a dramatic exit post. I didn't leave because of a fight with management, nor have I lost faith in Amazon or Aurora. In fact, I left on great terms, with the understanding that I might be politely asking for my badge back in 6 to 12 months.
So what will I be doing during this time? In short, I'll launch my own startup.
What Valuable Company Is Nobody Building?
Now, I don't agree with Peter Thiel on much of anything, except for his observation that startup ideas ought to answer the question, "What valuable company is nobody building?" My answer hinges on the following five arguments.
1. Local Models Are Now Good (Enough)
For the past few years, running a model locally mostly meant using Ollama to load a small model of a few billion parameters that was moderately useful for toy projects, but struggled with complex instructions, basic logic, or even precise output formatting. Well, it's pretty clear to me we're not in Kansas anymore.
For starters, people already report using local models like Gemma and Qwen as their daily drivers. Every time a new model is released, organizations like Unsloth publish quantizations on Hugging Face. There are now plenty of alternatives to Ollama, for example, Jan, LM Studio, oMLX, or Nativ. Finally, custom inference engines beyond llama.cpp, like DwarfStar, allow you to run quasi-frontier models like DeepSeek-V4-Flash on commodity hardware, including MacBooks.
2. MacBooks Run Local Models Quite Well
Modern LLMs are built on top of the attention mechanism, which can be proven to be memory bandwidth-bound at token generation time. MacBooks equipped with Apple Silicon chips in their Max variants feature a memory bandwidth of 400 GB/s or more. Comparable "system-in-a-box" products such as NVIDIA's DGX Spark and AMD's AI Halo offer 273 GB/s or less.
Inference-time scaling is a key ingredient of reasoning LLMs. Because intermediate outputs cannot be trusted, their execution must take place within a sandbox. Apple's Virtualization framework and Containerization package provide excellent primitives for that.
3. Companies Are Already Buying MacBooks
Powerful MacBooks are routinely issued to engineers and analysts; for instance, my last laptop at Amazon was an M4 Pro with 48 GB of RAM. Apple's Device Management service is widely deployed and well understood by enterprise IT departments. By contrast, alternatives for local inference, such as custom rigs with multiple GPUs, would require IT teams to go out of their way to support them.
4. Integration with Internal Tools Is a Moat
In Schlep Blindness, Paul Graham defines a "schlep" as a tedious, unpleasant task that people—especially hackers—avoid. In the local AI ecosystem, this manifests as an abundance of model harnesses, runners, and inference engines, which are fun engineering problems, while the unglamorous work is cast aside.
Some of that unglamorous work consists of connecting local models to messy internal infrastructure: legacy APIs, bespoke authentication schemes, and organically grown but undocumented database schemas. Doing this part well creates a far more defensible moat than doing what everyone else seems to be doing, which is rapidly commoditizing as a result.
5. It's Time to Cross the Chasm
In Crossing the Chasm, Geoffrey Moore describes the "chasm" as the wide gap between early adopters and the early majority. While the former are okay with half-baked products that are innovative first and useful later, the latter demand whole products that solve problems they've already identified but cannot solve on their own.
Local LLMs are technologically mature, but they have yet to cross this gap. The current end-user experience works for engineers, but we cannot expect a non-technical employee to browse Hugging Face, select the right model, and figure out which quantization fits their available RAM—all while navigating corporate IT policies that require manager approval to install any software.
Therefore, Air Gap
Which brings us to Air Gap, the startup I'm getting off the ground this year. In short: local LLMs on MacBooks for retail bankers. Why retail bankers in particular? I happen to live in Toronto's financial district, which gives me easy access to an industry overflowing with repetitive work, legacy systems, and strict privacy constraints. The customer conversations I've had so far have been nothing short of encouraging.
That's all for today: the idea is to post weekly updates every Monday, so (hopefully) I'll see you again on August 17th!