How to Build an AI Startup with Other People's Money

The AI labs are selling you $10,000 a month in computing power for $600. They're doing it at massive losses. And they have very specific reasons you should understand.

In this episode, I break down the $670B subsidy era fueling healthcare AI — what it means for builders, when it ends, and four strategies to exploit it before the economics correct themselves.

We cover:

• Why AI pricing follows the exact same playbook as early AWS
• The real numbers behind OpenAI's and Anthropic's losses
• Four strategies to maximize subsidized compute (from $600/mo subscriptions to $2,500 local hardware)
• Healthcare startups building real businesses on below-cost AI
• The subsidy timeline: when prices normalize and what to do before they do

Whether you're a healthcare founder, an operator evaluating AI tools, or just trying to understand why trillion-dollar companies are giving away compute — this one lays it out with data, not hype.

Watch on YouTube: https://youtu.be/uDB_VJAX05g

Full sources and the deep-dive subsidy timeline analysis: https://open.substack.com/pub/danmccoymd/p/the-golden-age-of-healthcare-ai-and?r=11z0su&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Transcript

SPEAKER_00 0:00

Imagine someone told you they'd invest $50,000 in your healthcare startup. No board seat, no equity negotiation, no pitch deck revisions. You'd assume they were running a scam. But that's essentially what's happening right now. Except the investors are anthropic, OpenAI, and Google, and they're not writing checks. They're selling you $10,000 a month in computing power for $600. They're doing it knowingly. They're doing it at massive losses, and they have a very specific reason for doing it that most founders aren't really thinking about. Here's a contrarian take that nobody in the AI hype cycle really wants to hear. We're living through the most subsidized period in the history of software development. And if you're building a healthcare AI company, like we did here at Rocket Tools, the next 12 to 18 months represents the arbitrage window that will not repeat. To understand why this matters, you need to understand what happened in the early days of cloud computing. Amazon Web Services launched in 2006 and priced its compute at rates that made no economic sense. They were losing money on every customer. The strategy was simple. Get developers addicted to the infrastructure, make switching painful, and then gradually raise prices once the ecosystem was locked in. It was a spectacular success. AWS went from a side project to generating $100 billion in annual revenue. By the time competitors caught up, Amazon had a decade-head start and switching costs that kept customers paying 30 to 40% premiums. The AI labs are running the exact same playbook, and they're doing it at a scale that makes Amazon's early cloud subsidies well look like a rounding error. In 2025 alone, four hyperscalers, we're talking Amazon, Alphabet, Meta, and Microsoft. They committed $670 billion in AI infrastructure spending. That's not a typo. $670 billion. Meanwhile, American consumers spend roughly $12 billion a year on direct AI services. The math doesn't work. It's not supposed to work. Well, not yet. The mainstream narrative goes something like this: AI tools are getting cheaper because the technology is improving. Moore's law for language models. Costs come down, access goes up, everyone benefits. That's a lovely story, but it's also very incomplete. Yes, inference costs have dropped dramatically. Epic AI tracked this, and in GPT-4 equivalent performance went from $20 per million tokens in late 2022 to 40 cents in 2025. That's a 50x decline. That's real and that's oppressive. But here's what the optimists leave out. Sam Altman, the CEO of OpenAI, went on X in January 2025 and he said, and I'm quoting him directly, insane thing. We're currently losing money on OpenAI Pro subscriptions. That's the $200 a month plan. He personally set the price. He thought they'd make money on it. And guess what? They didn't. OpenAI posted a $5 billion loss on $3.7 billion in revenue in 2024. Their inference costs hit $8.4 billion in 2025 and are projected to reach $14 billion in 2026. Cumulative cash burn through 2030. Analysts project $665 billion. Anthropic, which is our preferred model, isn't too much better. They brought in $4.2 billion in revenue while spending $7.2 billion, burning $3 billion in cash in 2025 alone. Profitability has been pushed back. Cash break-even is now projected for 2028. These companies aren't offering you cheap AI because they figured out the economics. They're offering you cheap AI because they haven't. Let me get specific because this is where the opportunity becomes almost absurd for healthcare builders. A Claude Max subscription runs $200 a month. OpenAI's Codecs and ChatGPD Pro another $200 for $400 combined. For roughly $600 a month, you have access to frontier class AI models that would cost $10,000 or more at AI rates for equivalent usage. Let that sink in. Martin Alderson did the math on Claude CodeMax users. That's somebody like me. The API equivalent spend for a heavy developer runs about $5,000 a month. Anthropic's actual compute cost on that roughly $500, meaning they're losing $100 to $300 per month on their most active subscribers. They know this, but they're doing it anyway. Why? Because model switching has friction. If your entire development stack, your prompts, your workflow, your team's muscle memory is built around clawed like ours, you're not switching to Gemini when prices go up. You'll pay the increase, and that's the bet. For healthcare startups, this means you can build clinical documentation tools, prior authorization systems, patient communication platforms, all running on models that are being subsidized by venture capital that won't flow forever. There's a deep analysis on the subsidy timeline. I'll put in the Substack post for that episode. But the short version is this analysts expect meaningful price normalization within 12 to 24 months. Here's something most healthcare founders don't realize. A $2,000 to $3,000 PC with a modern GPU can run open source models that rival what was state of the art 18 months ago. Llama, Mistral, Quinn, you may have heard of these. These models are free. The hardware is a one-time cost. Your ongoing expense is electricity. That's actually what we've done here at Rocket Tools. A $14 billion parameter model in 2026 matches what a 70 billion model could do in 2024 on coding and summarization benchmarks. You can run these at 60 to 120 tokens per second on a mid-range setup. For tasks like data structuring, clinical note parsing, appointment scheduling, you don't need a frontier model. You need a fast, cheap local one. And it could be free once you put the infrastructure in place. At Rocket Tools, we run lightweight models on our hardware just for these kinds of workloads. The initial setup cost is real, maybe $2,500 for the hardware, a weekend to configure it. But after that, your marginal cost per inference is essentially zero. Not every task needs clawed OPAS or GPT-4.0. Chimi's K2.5 model, DeepSeaks V3, Anthropics on Haiku, these are five to ten times cheaper than frontier models and perfectly capable for 80% of workloads. The healthcare startups I see wasting money are the ones running every API call through the most expensive model. The smart ones root intelligently, frontier models for complex clinical reasoning, lightweight models for everything else. For privacy-sensitive healthcare workloads where you can't use managed hosting, GPU cloud rentals have gotten remarkably cheap. An H-100 on Vastai runs about $1.87 an hour. An A100, still a serious machine, is under $1 an hour on spot markets. AWS cut H-100 pricing by 44% just last June. You can spin up a HIPAA compliant sandbox, run your model, process your data, and shut it down. No long-term commitment, no capital expenditure, a dollar an hour to run a model that costs millions to train. Now let me be honest about what this actually looks like in practice because the numbers get real very quickly. The healthcare AI market hit $39 billion in 2025 and is projected to reach $504 billion by 2032, a 44% compound annual growth rate. AI-enabled healthcare startups are raising at an 83% premium over their non-AI peers, according to Rock Health. The average speed round for healthcare AI is $4.6 million. But here's what's more interesting than the mega deals. Look at what lean teams are doing. Lunabill, a medical-billing AI startup, hit $764,000 in contracted annual recurring revenue since launching in July of 2025. They achieved 10x improvement in claims, followed up per biller in the first week, 100% conversions of pilots to paying customers, 50,000 automated calls. Attuned intelligence, founded by DeepMind alumnus, raised $13 million, went live in 10 days at Lowell Community Health Center, and now handles every mainline call 24-7, automating 70% of interactions. These aren't billion-dollar operations, these are small teams leveraging subsidized compute to build real products solving real problems. Billing, scheduling, documentation, the unsexy infrastructure of healthcare that everyone complains about and nobody really wants to fix manually. The ambient scribe segment alone generated $600 million in revenue in 2025, a 2.4x year-over-year increase. A bridge holds 30% of that market. They raise $300 million at a $5.3 billion valuation, but the market is still early enough that a two-person team with $600 a month in AI subscriptions can build a competitive product in a niche vertical. My confidence on how long this window stays open? Maybe 60%. The subsidies could last longer if the AI labs raise more capital. They could end faster if investor patient finally runs thin. Uptech Studio put it bluntly, when the price correction comes, it will likely be abrupt and not gradual. One analyst, Steve Smith, at our Dallas predicts agentic AF subscriptions will increase 10x to 100x from January 2026 levels by the end of 2027. I think that's aggressive, but the direction is probably right. If you're a healthcare entrepreneur or a builder like me, here's what this translates to in practical terms. First, your $600 a month AI budget, it's not a cost. It's subsidized RD. Anthropic, OpenAI, and Google are collectively spending billions of dollars to give you computing power below cost. Use it, build with it, prototype aggressively. The cost of experimentation has never been lower and it will not stay this low. Second, pick your infrastructure strategy now, not later. If you're building something that requires data privacy and in healthcare, most things do, get comfortable with local models or cloud GPU rentals today. When subscription prices normalize, you'll need alternatives already tested and already integrated. Third, the mode isn't the model, it's the workflow. Everyone has access to the same subsidized compute right now. The startups that win will be the ones that use this window to build proprietary data pipelines. They'll earn clinical trust and lock in customer relationships. When compute prices finally rise, customers won't leave if your product is embedded in their operations. Sound familiar? That's the AWS playbook again, except this time you're the one running it. Fourth, don't chase the meg around. 62% of digital health VC is going to AI-enabled startups. The money is available, but Lunabild didn't need $100 million to reach $764,000 in ARR. They needed a good product and subsidized compute. The bootstrapped healthcare AI company is a real category right now and it won't be there for much longer. Here's the uncomfortable truth about this golden age. The reason you can build a healthcare AI company for $600 a month is not because the technology is cheap. It's because some of the richest companies in history are subsidizing your compute bill to win a platform war. You are the beneficiary of a $670 billion bet that most of these companies will eventually lose. This is not a permanent state of affairs. It's an arbitrage window created by venture capital and patience and infrastructure arms race. The founders who recognize this, who build real products with real customers on subsidized infrastructure, will have functioning businesses when the subsidies end. Everyone else will have a prototype and an AWS bill they can't afford. The money is free right now. It won't be for long. If you found this useful, hit subscribe. Sources and additional analysis on the subsidy timeline are in my Substack. Links in the description.