He runs the company alone, charges a flat subscription, and pays for AI by the token. That combination quietly turns his best customers into his biggest losses. Here is the margin math, and why open-weight Chinese models became the escape hatch.
Ben Broca built a $10 million-a-year business with no employees, then found he was losing money on many customer accounts because he paid Anthropic per token while charging a flat rate. He switched to free open-weight models from China. His problem is structural, not personal: usage-based AI costs running against flat subscription pricing invert the margin model that made software the best business in the world, and the heaviest users cost the most to serve.
Ben Broca sells AI tools to entrepreneurs, runs the whole company from his living room in Sausalito, and answers to nobody. He told the Wall Street Journal that “compromises make lukewarm results.” His own gross margin made one compromise for him anyway. To keep serving 10,000 paying customers without bleeding cash, he dropped Anthropic’s Claude and moved his workloads onto free open-weight models out of China.
What Happened
The Wall Street Journal profiled Broca on July 29 as the face of a new cohort: solo operators running seven- and eight-figure businesses with no staff. Stripe’s own data backs the trend. The number of solopreneurs on its platform clearing $1 million in revenue doubled between 2023 and 2025, and the count crossing $10 million nearly tripled. We covered the structural version of this shift in The Rise of the One-Person Company, where the labor never disappears, it moves off payroll and onto invoices, software, and compute.
Buried in the WSJ piece sits the detail that matters more than the headcount. Broca said he was losing money on many customers’ accounts while paying to access Claude, because Anthropic and its rivals bill by usage. He has since switched to free open-source models from China. The one-man company was underwater on unit economics, and the fix was a model swap.
The Backstory
For twenty years, software was the best business model anyone had found because of one number: gross margin above 80%. Once you built the product, serving the next customer cost close to nothing. Bandwidth and storage were rounding errors. That near-zero marginal cost is why SaaS founders could charge a flat monthly fee and watch margins widen as they grew.
AI-native software breaks that number. Every request a customer makes fires off tokens to a model provider who charges for them. The cost of serving a customer now scales with how hard that customer uses the product. Anthropic and OpenAI pushed the pain further downstream when they moved buyers from flat plans toward token-based billing, a shift we examined in the margin-migration piece.
Stack a usage-based cost under a flat subscription and you get an inverted P&L. A light user pays $20 and costs you $2. A power user pays the same $20 and costs you $40. Grow the base and you scale the loss-makers right alongside the profitable ones. That is the trap Broca fell into, and it is waiting for every founder who wraps a frontier model in a fixed-price product.
The Plan
Founders staring at that math have three levers.
Lever one: meter the pricing. Pass token costs through with usage-based tiers so heavy users pay for the compute they burn. This protects margin but scares off customers who want a predictable bill.
Lever two: switch to a cheaper model. Same product, smaller invoice per request.
Lever three: self-host open weights. Chinese labs publish DeepSeek, Qwen, Kimi, and GLM under permissive MIT and Apache licenses, which means a founder can run the model on rented GPUs and pay for hardware instead of paying per token. DeepSeek’s business model is built on exactly that giveaway: free weights drive adoption, paid API and cloud tiers earn the money.
Broca reached for levers two and three. The pricing stayed flat for his customers, and the model underneath got radically cheaper.
The Business Model Angle
The size of the prize is the story. Published output-token prices in mid-2026 put US frontier models in a different universe from the Chinese open-weight tier.
Claude Opus 4.8 lists at $25 per million output tokens. DeepSeek V4 Flash lists at $0.28. That is a roughly 89x gap on the exact line item that was sinking Broca’s accounts. Even the mid-tier comparison, a capable open model like DeepSeek V4 Pro at $0.87 against a Western flagship at $25, runs 25 to 30 times cheaper. Move a generation-heavy workload from the first group to the second and a customer who cost $40 to serve now costs a dollar or two.
Here is the part founders should sit with. In AI-native software, gross margin is a model-selection decision you make every quarter, not a property of the business you built once. The cheapest credible model wins the cost line. That reframes the whole category. A SaaS founder in 2015 could not cut cost of goods by 90% with one configuration change. An AI-native founder in 2026 can, by swapping a base URL.
The same fact that saves Broca also strips his advantage. If margin comes from picking a cheap model, and any competitor can pick the same model the same afternoon, that margin defends nothing. Value has to sit somewhere the model does not sell: the distribution he already owns, the proprietary data he feeds the model, the workflow a customer cannot rip out without pain.
The Risk
Cheap models carry a bill of their own, paid in different currencies.
Quality and consistency slip. Many hosts quantize open weights to fp8 to cut their own costs, which drifts output away from the released model. What benchmarks at parity in a lab can wobble in production.
The bill does not always fall. When token prices drop, usage tends to climb faster than the price, so total spend can rise even as the per-token rate collapses. Enterprise finance teams learned this the hard way, a dynamic we traced in the margin-migration analysis. A cheaper model is not a cheaper bill if your customers respond by hammering it harder.
Geopolitics and compliance bite. Chinese models raise data-residency questions, enterprise procurement teams block them, and DeepSeek has drawn criticism for routing user data to China. Export-control policy swings both ways and fast, as the mid-2026 suspension and restoration of certain frontier models showed, so any founder betting the company on a single foreign provider is pricing a policy risk they do not control.
Dependence does not disappear, it changes address. Broca traded reliance on Anthropic for reliance on a Chinese lab or a third-party host. Free weights stop being free the moment you use a provider’s API, and providers ration access when demand spikes. Moonshot did exactly that, capping access to a popular Kimi model after it got too popular, which we covered in the China AI price-war piece. Self-hosting removes the vendor but hands you a GPU bill and an ops burden that can rival the salaries the one-person model was supposed to avoid.
Quick Questions
How does a $10 million company lose money on customers? It charges a flat subscription while paying for AI by usage. Heavy users generate more token cost than their fixed fee covers, so growth in that segment scales losses instead of profit.
How much cheaper are Chinese open-weight models? On published mid-2026 output pricing, roughly 5 to 30 times cheaper than US frontier flagships across the tier, and close to 90 times cheaper at the extreme of Claude Opus against DeepSeek V4 Flash.
Are they as good? For many production tasks, close enough, and several lead specific coding and agent benchmarks. Quality depends heavily on the host and whether the weights are quantized.
Does a cheaper model mean a cheaper bill? Not always. Cheaper tokens invite heavier use, and total spend can rise even as the per-token price falls.
Can an AI-native business hold a durable moat? Not from model choice alone, since competitors can copy it instantly. The defensible margin comes from distribution, proprietary data, and workflow lock-in.
The Business Model Analyst Take
AI-native software breaks the rule that made software the best business model in history. Marginal cost is no longer close to zero. It is a metered invoice from a model provider, and it lands every time a customer clicks. That one change moves gross margin from a fixed property of the business to a procurement decision the founder revisits constantly, where the cheapest credible model wins the line item.
Broca’s move to Chinese open weights is the rational read of that math, and his cohort will follow him rather than lead. The catch is the same fact seen from the other side. A margin you buy by swapping vendors is a margin any competitor buys the same afternoon. The model is now the cost line, not the product. The founders who keep their margins will be the ones who own something the model does not sell, the audience they built, the data they hoard, the workflow the customer cannot leave. Pick the cheap model and win the quarter. Own the distribution and win the decade.
