The Work Behind My $100M Goal

Why I’m doing this

AI is the biggest opportunity I’ve seen in 12 years of business experience. I’ve spent about 5,000 hours working with it over the past year, much of that time building products and learning the technology underneath them.

My goal is $100M. I come from business: understanding what people need, creating value and making money. This past year has given me a much deeper understanding of the engineering behind a product. I’m bringing those two kinds of knowledge together.

I’m keeping the projects private for now, but I can share the work behind them.

My year with AI

I tried hundreds of things, often staying with the same project for two or three days. My workdays now often run 18 to 20 hours, with very few outside interruptions. I found work I love, and that explains many of the decisions I’ve made about how I spend my time.

I had already taken a break from Instagram and Facebook. Once I got absorbed in AI, I stopped wanting to return. For about a year, I barely went out or reached out to people. I no longer use social media, a mobile phone or ordinary phone calls.

I write most of my code with coding agents. Across my AI work, I’ve measured around one trillion tokens in total. As of October 2026, I’m using around 10 billion tokens per day.

For perspective, AT&T is a telecommunications giant whose stock-market value was around $200 billion in February 2026. Its reported AI usage grew from roughly 8 billion tokens per day in 2025 to 45 billion per day in July 2026. I’m not AT&T. I’m one person doing the building on these projects. These are different workloads, not equivalent business output, but the comparison gives some perspective on the scale of AI an individual can now put to work. To me, those numbers show the scale of AI an individual can now put to work.

In June 2026, Sam Altman said OpenAI’s leading token user consumed about 100 billion tokens a month. At my current daily rate, 30 days would add up to around 300 billion tokens.

I’m not an engineer at OpenAI or Anthropic with a major AI lab behind me, either. Next to my own budget, the resources of those labs and a company like AT&T feel almost limitless. I’m paying for my own experiments and building my own setup.

My guess is that I use more tokens per day than 99% of individual engineers at OpenAI and Anthropic, and that around 500 people worldwide combine this scale of independent AI usage with workdays like mine.

In August 2026, 4.5% of eligible consumers in YipitData’s US e-receipt panel had an active paid personal subscription to ChatGPT, Gemini or Claude, according to a16z’s October report. My guess is that 0.5% of people in the US personally spend around $200 a month on AI. I’d put the worldwide number of customers spending $2,000 or more a month at roughly 50,000, counting individuals and companies together.

Separately, Bank of America’s March 2026 report found that about 3% of its customer households paid for AI services, with a median monthly spend of $20 in February.

What I had to learn

That work took me into code, design, repositories, Docker containers, servers and system architecture. I wanted to understand how the pieces of a complicated project fit together, what could break and how to find the problem.

I invested around $200,000 in learning, experiments and infrastructure. My setup includes around 100 devices, with five Macs, an NVIDIA-equipped computer and local servers. I wanted to understand what happens on machines I run myself, not only use a service someone else built.

Every layer under an AI product affects what it costs and how it fails. At data-center scale, power and cooling constrain the hardware. On my own machines, memory capacity and bandwidth help determine which models I can run and how they perform. The model itself is only part of the memory requirement. The KV cache stores intermediate calculations from earlier tokens so the model can reuse them. Longer conversations and more requests running at once generally require more memory. A model fitting on a machine does not mean that machine can serve a busy product.

Docker makes software easier to move between machines, but it does not remove every difference between them. GPU software still needs compatible drivers. A successful local run is the start of deployment, not the end.

I also had to understand the cost of running models. Long agent runs can repeatedly send large amounts of context. Caching, model selection and limits on a task’s steps or spending all matter. AT&T reports that its efficiency improvements have reduced AI costs by as much as 90%, saving millions of dollars. My resources are very different, but I have to think carefully about those costs too. Local machines can make sense for private experiments and batch jobs; customer-facing services need infrastructure chosen for dependable operation.

Distributed systems made me think differently about failure. A payment request can time out even though the payment succeeded. With a payment API that supports idempotency, retrying the same request with the same key helps prevent a duplicate charge. Clear agreements about the messages services exchange, and careful retry rules, matter as much as the code inside each service.

I learned to plan for interruption and recovery. Long-running jobs need checkpoints so they can resume from recently saved progress. Customer records need backups that have actually been tested by restoring them. Database changes need a migration plan, and releases need gradual rollouts and a recovery plan. Logs and traces help explain what happened when something fails.

AI adds another layer of uncertainty. A model can produce a plausible answer that fails a real test. Prompts and model changes need evaluation. Agents need narrow permissions and human approval before taking consequential actions such as moving money. An imported document can contain instructions designed to manipulate the model; reading that document should not give it control of the application or access to private data.

Why a small team can do more now

Software feels much more fluid to me now. I can change a design, rebuild part of a product and test another approach much faster. It felt like becoming a hundred times more productive.

With AI, someone can start building useful software while learning the fundamentals. Running a dependable product is a further step.

Think about an eighteen-year-old with a laptop and access to good AI models. They don’t need my setup or anything close to my budget to begin. They can start with a task they understand, use AI for part of it and check the result themselves. They can learn a lot before spending much money.

The tools are improving too. Google DeepMind’s AlphaEvolve has helped improve algorithms used in computing and AI training. My own bet is that the next five to ten years will bring an extraordinary amount of technological change. But the reason I’m building now is what I can already do today.

To me, this is a once-in-a-lifetime opportunity. It doesn’t remove the need to understand customers, deliver good work or manage a business. It gives me a better way to put those skills to use.

Choosing what to build

Business taught me that working hard and working on the right thing are different skills. I’ve had to learn both, along with knowing what to leave alone. AI gives me more options, but it can’t choose the business for me.

I’m putting that ability to work on the businesses I believe in. That is the work behind my $100M goal.