Building a Multi-Headed Business from a YouTube Channel

Spencer’s path into his current work began with a YouTube channel. In 2022 he launched Start a Pack after earlier experience running Clean Technology, a company he started in 2009 that produced the Clean Router. That earlier venture went from zero to roughly a million in annual recurring revenue and later back down near zero, giving him a range of stories about growth and setbacks. He formed an LLC around the YouTube effort and committed to publishing five to six videos a week. Four years later the channel has passed 9,000 videos.

Subscriber counts sit near 24,000, though he notes that subscribers function more as a vanity metric. The actual monthly audience of people who watch the videos is about 130,000, producing five to seven million video views a year. A full-time video editor now supports the effort: Spencer records roughly five live sessions each week, and the editor turns that material into both long-form videos and short-form clips, resulting in 50 to 60 published pieces weekly. The channel drives business to both the consulting practice and the products.

The overall operation has grown into what Spencer describes as a three-headed (or even 400-headed) structure. It includes the YouTube channel, a licensed vocational school that operates as a registered apprentice program, software consulting, and the development of original products. The school hires people with no prior experience, trains them as software developers, and moves them into consulting roles. Because the team already builds software for clients, it made sense to start building internal products as well. Start a Pack originated as the YouTube channel name and became the umbrella for these activities.

AI’s Effect on How Software Gets Built

AI is altering nearly every part of the business, from customer interactions to the process of writing software itself. Spencer’s setup includes AI-enabled machines in an R&D lab and a private data center that supports ongoing work. He had already spent about half a decade around machine learning, helping produce data for ML teams in earlier roles, so the direction of the technology was not entirely unexpected. The GPT-3 moment still surprised most people, including him, and prompted the same scramble to keep up that many other developers experienced.

One concrete outcome is an open-source project called Open Mono Agent AI. It functions as a full harness that lets anyone download and run a complete AI agent stack. The software can install on a small device and operate a full local AI environment. It began as something closer to a Cloud Code replacement and evolved into a broader tool that also competes with projects like Open Claw. Users can run models, add webhooks, and own the entire engine. Spencer encourages people to try it simply to understand how AI actually works, regardless of whether they adopt the project for production use. He sees frequent consulting inquiries from executives who know they need to “do AI” but have little clarity on what that means in practice.

Real-World Pros, Cons, and Limits of Building with AI

Code can now be produced very quickly, so raw writing speed is no longer the primary constraint. The harder work is making large codebases and multiple pieces fit together reliably. Spencer hears frequent claims that software and SaaS are dying, yet he expects specialized domain knowledge around existing SaaS products to remain valuable. Large platforms such as Salesforce may face pressure, but they are unlikely to disappear; Claude has already partnered with some of them. For smaller shops the ability to generate code fast is real, yet several caveats remain.

Security has moved to the front of the priority list. It was always important, but failing to address it at the outset now creates immediate risk. There is also a wide gap between public hype and enterprise reality. Technical founders may talk about “vibe coding” something quickly, yet real organizations generally do not want to own and maintain a collection of hastily assembled tools. Spencer’s own company recently moved from Zoho CRM to HubSpot. They had previously built internal tools and could have chosen open-source options, but HubSpot offered stronger integrations, MCP capabilities, and connections to other software they needed. The monthly cost of two or three thousand dollars was preferable to the ongoing burden of hiring and retaining someone to support a custom system that only needs occasional changes.

He compares building custom software to adopting a puppy: most companies prefer to buy solutions that remove problems rather than take on the long-term responsibility of raising and maintaining something themselves. Building internally makes sense mainly when it creates genuine intellectual-property value or delivers substantial cost savings—for instance, if an organization is already spending half a million dollars a year on a platform like Salesforce.

One product the team is developing, Swiftcase Legal.AI, illustrates these points. Many lawyers are already using Claude heavily, yet Claude does not manage an entire firm’s workflow. Security and specialization matter: client data cannot simply be sent to a frontier model. Roughly 90 percent of law firms have fewer than ten lawyers. Larger firms may be served by companies such as Harvey, but smaller practices currently have fewer options. Swiftcase Legal.AI is being built as a SaaS platform with an embedded AI workflow, hosted in a private data center so that no client data touches ChatGPT, Claude, or similar services. This approach addresses concerns from firms handling sensitive matters, such as cases involving government entities. The GPT-3 era normalized the idea of sending large volumes of data to cloud models with little regard for privacy; many regulated industries still refuse to accept that practice.

AI has also become an effective lead generator. People prototype ideas with tools like Claude, Claude Co-work, or scheduled tasks. Eventually the prototype outgrows what they can maintain, or they need deeper functionality, and they seek either a developer or a more complete product. The same “puppy” problem appears: data integrity, security, and ongoing support quickly become larger than expected. A common joke after the GPT-3 launch captured the gap—someone claims they built a full website with the model, only to reveal it still runs only on localhost.

Data Privacy, Local Hardware, and Control

Spencer runs local models primarily for data sovereignty and control. He will not send customer data to the cloud, regardless of stated privacy policies. Frontier models train on the data they receive, including material from unpaid or private accounts. Running models locally also avoids vendor lock-in. When OpenAI removed the 4o model, products that depended on it collapsed. With local infrastructure he can swap models as needed and match a specific model to a specific problem rather than being limited to one provider’s options.

The hardware is deliberately ordinary. The highest-end card in use is a 5090; list price is around three thousand dollars, though real-world prices have reached roughly eight thousand. A 3090 or 3090 Ti that once cost eight hundred to a thousand dollars now runs fifteen hundred to two thousand dollars, yet still supports impressive work. Many of the products are being built on cards of this size or smaller. Older hardware remains useful: a six-year-old GPU and even eight-year-old V100 cards are running 90-billion-parameter models. Smaller models are being pushed onto cards such as a 3060 with 12 GB of memory. Tools like LM Studio are used, and Open Mono Agent AI provides the surrounding harness so that webhooks and custom logic can be added while the operator retains full ownership.

Cloud GPU rental (for example through providers that offer transient capacity) has been considered, especially where the provider guarantees that data is not used for training. For now the preference is to host everything directly. If scale demands it, that stance could change, but the architecture already separates shared instances from dedicated corporate servers. One client—a contract law firm working with three sheriff’s departments—cannot host in the cloud under federal regulations and therefore runs on its own box. Data centers themselves represent concentrated points of failure; examples from the Ukraine-Russia conflict and an incident in Bahrain illustrate the risk. Over time, some organizations may move further away from centralized cloud facilities for both reliability and privacy reasons.

Spencer expects a wave of legal challenges over the next five years focused on data that has been used for training by OpenAI and Anthropic. Microsoft has carved out a clearer position with government contracts that specify hosting details. In his view, the frontier labs have prioritized growth and operated on a “forgiveness rather than permission” basis; he sees no reason to expect different behavior around training data.

Pace of Change, Adoption Numbers, and Practical Advice

The hardest part of building software or a software company right now is the speed of change. Spencer has worked in software since 1998 and has never seen a period like the last three years. Even people with longer careers struggle to forecast five years ahead. New frontier-level models that once appeared monthly now arrive weekly or daily. A model such as Qwen 3.8 27B, only two months old at the time of the conversation, already outperformed the frontier models of a year earlier by nearly 40 percent and can run on essentially any decent graphics card. At a certain point the models feel “good enough,” and differences begin to resemble the gap between iPhone and Android—noticeable preferences and ecosystems remain, yet core capabilities converge. Models are becoming more like a commodity calculator.

One intriguing development is a company that converted a model into a fixed hardware chip capable of running roughly a hundred times faster. The model itself cannot be updated after fabrication, but the approach suggests future systems may include a CPU, a GPU, and a dedicated coprocessor holding a “good enough” model that handles the majority of everyday tasks. Similar specialization is already visible in Tesla’s AI4 chips and the planned AI5 and AI6 chips for vehicles and Optimus. Hardware acceleration of this kind should eventually lower costs, even if the models themselves cannot be patched like software.

Current adoption remains low. Only 1.2 percent of American households that contain at least one person currently pay for an AI service, and those users are essentially all developers. By comparison, about 70 percent of America pays for a Netflix subscription. Reaching even 10 percent saturation would represent a major jump, which helps explain why hyperscalers continue to build data-center capacity aggressively. Whether most people will ever pay is an open question. Many users, including family members of both speakers, rely on free tiers such as Google Gemini. Over time the subsidies that support free access are likely to shrink, free usage will be capped more tightly, and prices will rise once a critical mass of users is hooked—the same pattern Netflix followed after years at a low introductory price. Upcoming IPOs for Anthropic and OpenAI are expected to mark the point at which the current “gravy train” of heavy subsidies begins to end.

The practical recommendation is straightforward: just start using AI. The advice mirrors what Spencer has told people about coding for thirty years—there is no perfect entry point, so begin doing something. Download LM Studio and experiment, or try the free open-source Openmonoagent.ai project to see the full stack. The goal is simply to move into the small group already paying for and learning AI rather than remaining in the 98.8 percent who have not.

People interested in custom software work or the broader activities around Start a Pack can visit startuphakk.com. The combination of rapid code generation, heightened security requirements, persistent domain expertise, and still-limited consumer adoption continues to shape both the opportunities and the constraints for founders working at the intersection of consulting and product development.