Daniel sent us this one, and it's a bit different — he's basically opening the books on the show itself. We've got about a hundred and eighty thousand plays, zero sponsors, and operating costs running around two hundred dollars a month, almost all of it going to Modal for serverless GPU inference. The question is: who should we actually pitch for sponsorship when the time comes, what kind of ad formats could work without degrading the show, and are there non-profits whose mission aligns closely enough that they might underwrite the whole thing? And then there's the expansion question — multilingual episodes, video — whether that's worth doing and how to fund it. So this is basically a strategy session on how to keep a niche AI-generated educational podcast alive without selling its soul.
I love that we're doing this transparently. Let's start by getting the numbers straight. Here's exactly what this show costs and what we're getting for it. The current burn rate is about two hundred dollars a month. Ninety-five percent of that is Modal — serverless GPU instances handling the TTS inference, generating the voices you're hearing right now. The remaining five percent is split between the Deepseek API for script generation and Fal for cover art. That's it. Twenty-four hundred dollars a year. For a library of educational content that's getting a hundred and eighty thousand plays.
Which, by the way, we should acknowledge: that play count isn't audited yet. No Podtrac, no Chartable, no third-party verification. Industry standard for unverified podcast feeds is somewhere between fifteen and thirty percent bot traffic. So the real human number is probably somewhere between a hundred twenty-five and a hundred fifty thousand. Still substantial for a show that covers how gyroscopes work one week and diplomatic protocol the next.
And that's the thing about the long tail. Episodes on niche topics don't spike and die — they accumulate. Someone searches "how does a sundial actually work" three years from now, they find our episode, they listen. That's the asset we're building. But let's address the first misconception most people have: a hundred and eighty thousand plays sounds small to someone who thinks in YouTube-scale numbers, but in podcasting, especially niche educational podcasting, that's a real audience. At a twenty dollar CPM — which is conservative for a tech-literate, educated listener base — that's about thirty-six hundred dollars in theoretical ad revenue per hundred and eighty thousand plays. The issue isn't that the audience is too small. It's that we haven't monetized it at all.
CPM, for anyone who doesn't live in Herman's spreadsheet brain, is cost per mille — cost per thousand listens. Twenty dollars per thousand is a niche-education rate. Mass-market comedy podcasts might get fifteen. True crime gets higher because the audiences are huge and loyal. But the key variable here isn't raw size — it's engagement and demographic precision. A podcast where listeners are AI developers, engineers, educators, and curious generalists is worth more per listener than a podcast where listeners are just...
And that brings us to the obvious question: with those numbers in hand, who would actually want to pay for any of this? Let's build a pitch list.
Before you do — I want to flag something about the mission statement, because it's going to matter for who we approach. The show's purpose, as Daniel framed it, is to empower individuals with robustly verified AI-generated content that explains how the world works, without regard to how well-trodden a topic is. That's a public-good framing. It's not "we make entertaining content about AI." It's "we use AI to make educational content accessible." That distinction changes who might care.
It really does. Okay, pitch list. I've got seven candidates, and I'll go through the why for each. First and most obvious: Modal themselves. They've already given a grant, which means they see value in what we're doing. A deeper partnership — something like a "Powered by Modal" badge in show notes, or even better, a case study they can publish about how a podcast runs TTS inference on their serverless GPU infrastructure — that's marketing for them. They get to say "this entire podcast is generated on Modal." We get credits or direct cost offset. The alignment is basically perfect.
It's genuine. We're not inventing a use case for them. We are the use case. The pitch writes itself: "You already know us. We're the weirdest thing running on your infrastructure. Want to make it official?
We use their API directly for script generation — not through a third party, not through OpenRouter, directly. That makes us a reference customer. An affiliate model could work here: "This episode was scripted using Deepseek" with a link and maybe a small credit per signup. Or just a usage credit arrangement. They're competing in a crowded market — having a production podcast that publicly uses their model is worth something.
Strong alignment with open-source AI, educational content, and community. They sponsor a lot of things — the Hugging Face paper reading groups, community events. A podcast that's essentially an AI-generated educational library fits their ethos. And they've got a platform to promote it on. The pitch is: we're building exactly the kind of accessible AI education you advocate for, and we're doing it with open-weight models where possible.
They'd also probably appreciate the transparency of an episode like this one — pulling back the curtain on the economics.
Fourth: Weights and Biases. They do MLOps monitoring for machine learning pipelines. Our audience overlaps heavily with their customer base — people who train models, run inference, care about pipeline observability. A "Tool of the Week" segment where we actually use W&B to track something about the episode generation process would be genuinely interesting to listeners. It wouldn't feel like an ad — it would feel like behind-the-scenes content.
That's the key, isn't it? The ad has to be content. If the sponsor integration teaches the listener something about how the show works, it's not an interruption — it's part of the product.
Fifth: Replicate or Fal. We already use Fal for cover art. Replicate is similar — cloud-based model inference. These are small, scrappy companies that understand the indie AI ecosystem. A cross-promotion where they feature the podcast in their community and we mention them as our image generation partner — that's low-effort, high-authenticity.
A cloud provider like DigitalOcean or Linode. I know we don't use them for the main pipeline — Modal handles the GPU side — but the audience of indie developers and technical founders overlaps heavily. These companies have a long history of sponsoring developer podcasts. The Changelog podcast has been sponsored by Linode for years with a simple "Presented by" model — one mention at the top, one at the bottom, no mid-roll. It's tasteful, it works, and it builds brand association over time.
For a company like DigitalOcean, sponsoring a niche educational podcast is basically rounding error on their marketing budget. They spend millions on developer advocacy. A couple hundred dollars a month to reach a hundred thousand-plus technically literate listeners is a steal.
Seventh: an AI ethics or education non-profit. I'm thinking The Partnership on AI, or The AI Education Project. These organizations have mandates around AI literacy and public understanding. A podcast that generates verified educational content about how the world works — including AI topics — is literally doing the thing they exist to promote. They might not have large sponsorship budgets, but they have networks and credibility.
That's a good bridge into the non-profit conversation later. But before we leave the sponsor list — let's talk about the ad formats themselves, because that's where things usually go wrong.
Daniel was explicit: no heavy sponsor messages, no degradation of the content stream. So I've been thinking about three formats that could work. Format A: "Presented by" sponsorship. A single mention at the top of the episode — "This episode is presented by Modal" — and a single mention at the bottom. That's it. No mid-roll. No host-read ad copy about features and pricing. Just a quiet brand association. This is what The Changelog does with Linode, and it's been their model for years. Listeners barely register it as an ad — it's more like a publishing imprint.
The New Yorker of podcast ads. "This episode is supported by the John D. and Catherine T. " You hear it, you move on, the content is untouched.
Format B: "Tool of the Week" segment. Sixty seconds, naturally integrated, where the sponsor's tool is actually used to generate or analyze something in the episode. For example: "This episode's cover art was generated by Fal, using their Flux model. We prompted it with X and got Y — here's what we learned about prompting for cover art in the process." It's a mini-educational moment that happens to feature a sponsor. The key is that it has to be interesting. If we can't make it interesting, we don't run it.
"Listener Discount" — a promo code mentioned once, with a link in show notes. No host-read pitch about how amazing the product is. Just: "Listeners can get twenty percent off their first month of Weights and Biases at wandb.com slash weird prompts." Ten seconds, done. The value to the sponsor is the conversion tracking, not the airtime.
I think the principle underlying all three is the same: the ad must never be longer than it is interesting. If a sponsor wants a sixty-second host-read where we pretend to be excited about Kubernetes monitoring, the answer is no. If they want us to actually use their tool to do something cool and talk about what we learned, that's content.
That's the pitch to sponsors, really. We're not selling access to an audience that can be interrupted. We're selling association with a show that smart people trust. The value isn't the number of ears — it's the quality of the attention.
Which brings us to the second path. Sponsors are one route. But there's another that might be even more aligned with the mission Daniel described: non-profit underwriting.
This is where things get interesting, because the numbers are so small relative to what non-profits spend on communications and public education. Let me walk through the candidates. First: The Internet Archive. Their mission is universal access to knowledge. They archive websites, books, software, audio — they're building the Library of Alexandria for the digital age. A podcast that generates a permanent, growing library of verified educational audio content about how the world works — that's a modern extension of what they do. The pitch is straightforward: for twenty-four hundred dollars a year, you help ensure this educational archive continues to grow and remains freely accessible. In exchange, we include a "Supported by the Internet Archive" credit and a link. No editorial control. No content review.
The Internet Archive gets something tangible — a demonstration that their model of knowledge preservation extends to AI-generated educational media. It's a proof of concept for them too.
Second: The Wikimedia Foundation. The show's mission — "explain how the world works without regard to topic popularity" — mirrors Wikipedia's ethos almost exactly. Wikipedia covers everything from quantum chromodynamics to the history of the paperclip. We do the same, just in audio form, with AI-generated scripts that are verified for accuracy. Wikimedia has an annual budget in the tens of millions. A twenty-four hundred dollar annual grant is, I don't know, the catering budget for one board meeting.
The challenge with Wikimedia is they're very protective of their brand. They'd want to know that the content meets their standards for neutrality and verifiability. Which, honestly, is a good conversation to have — it would push us to formalize our verification process.
Third: The Mozilla Foundation. Focus on open web, digital literacy, trustworthy AI. They've been running the "Internet Health Report" and advocating for ethical AI development. A podcast that uses AI to create educational content, transparently, with clear disclosure about how it's made — that's a case study in responsible AI deployment. Mozilla could fund specific episodes or series on AI literacy topics, or simply provide general operating support.
Mozilla has a history of funding weird, experimental media projects. They're not just about Firefox. The foundation side does a lot of grant-making.
Fourth: The AI Now Institute. They're an academic research group focused on AI's societal impacts. They produce papers on algorithmic bias, labor displacement, AI governance. They might fund episodes that translate their research into accessible audio — essentially, commissioning educational content that explains their findings to a broader audience. The pitch is: you produce dense, important research that most people will never read. We can turn it into something people actually listen to.
That's a compelling offer. Academic research has a distribution problem. Podcasts solve distribution. If AI Now has a paper on, say, facial recognition in policing, an episode that explains it clearly and accurately is public education — which is part of their mission.
Fifth: The Raspberry Pi Foundation. Their focus is educational technology for underserved communities. They make low-cost computers, develop curricula, and promote digital literacy globally. The multilingual expansion angle is key here. If we're creating educational audio content in Spanish, Arabic, Hindi — that's directly aligned with their mission of making computing education accessible. Their annual budget is about thirty million pounds. A twenty-four hundred dollar sponsorship is zero point zero zero eight percent of that. It's a rounding error with outsized mission alignment.
The multilingual point is worth dwelling on. Let's talk about what expansion actually costs, because I think there's a misconception that going multilingual or adding video is prohibitively expensive.
It's really not — at least not for audio. Multilingual TTS is the main cost driver. ElevenLabs, for example, supports something like thirty languages now with reasonably natural voices. Adding one language — say, Spanish — would add roughly fifty to a hundred dollars per month in TTS costs, depending on episode length and voice quality. Spanish alone opens up about fifty million potential listeners in the United States, plus the entire Spanish-speaking world. The script generation cost is essentially the same — Deepseek handles translation or generates directly in the target language.
For roughly a hundred and fifty dollars a month, you could test Spanish-language episodes for three months, measure the increase in plays, and decide whether to expand further. That's a very cheap experiment.
The ROI case is compelling. Even a twenty percent increase in total plays from Spanish-language content would mean thirty-six thousand additional plays per year. At niche educational CPMs, that's seven hundred dollars in theoretical ad value — which more than covers the additional TTS cost. The economics actually work.
What about video? That's the one where I think the enthusiasm needs to be tempered.
I agree, and here's why. For an audio-first educational podcast, video adds marginal value unless it includes visual demonstrations — diagrams, code walkthroughs, animations. A waveform visualization or a static image over audio is what most "podcast videos" are, and they don't drive meaningful discovery on YouTube. YouTube's algorithm rewards watch time and engagement — a static image with audio doesn't perform well against actual video content. The production cost isn't the issue — OBS plus FFmpeg can generate waveform videos for pennies. The real cost is time. Editing, uploading, managing a YouTube channel, responding to comments — that's hours per week that could be spent improving the audio content or expanding languages.
Video is a time sink with uncertain returns for an audio-first product. The exception would be if a sponsor specifically wanted video placement and was willing to fund the production. But as a growth strategy on its own, it's probably not worth it until the audio side is fully optimized.
There's another factor. The long tail of niche topics — "how does a gyroscope work," "what is the history of the pencil" — these perform better in search-driven audio discovery than in YouTube's recommendation algorithm. YouTube wants content that generates engagement signals. A deep explainer on a niche topic might get steady but low engagement, which hurts it in the algorithm. Whereas in podcast apps, search-driven discovery means someone types "gyroscope explained" into Pocket Casts or Spotify and finds our episode. That's a perfect match. The platform matters for the content strategy.
We've got a list of potential partners. But before anyone starts sending emails, let's talk about what we should actually do first.
Actionable step one: implement auditable analytics. Podtrac or Chartable — both are free for basic tier, both provide IAB-certified measurement. Verified numbers are worth two to three times more in sponsorship negotiations because the buyer knows they're not paying for bot traffic. This should happen before any pitch goes out. You don't want to be in a conversation with a potential sponsor and have to say "well, we think it's about a hundred and eighty thousand, but we haven't verified.
It's not hard to set up. A redirect through Podtrac's prefix URL and you've got certified numbers within a month. The sooner that's running, the sooner we have real data.
Step two: start with a "Presented by" sponsorship from Modal or Deepseek. These are the most natural fits — we already use their products, they already know who we are, and the alignment is obvious. The pitch to Modal is: "We're a living case study of serverless GPU inference for content generation. Let us put your name on it." The pitch to Deepseek is: "Every episode is scripted on your platform. We'd like to make that a feature, not just a fact.
The advantage of starting with an existing relationship is that you can iterate on the format without the pressure of a cold pitch. If the "Presented by" mention feels awkward the first time, you adjust. By the time you're pitching cold to Hugging Face or Weights and Biases, you've got a working template.
Step three: approach two or three non-profits with a concrete proposal. Internet Archive, Mozilla, maybe Wikimedia. The proposal is specific: twenty-four hundred dollars per year for a "Supported by" credit and a link in show notes. Emphasize the educational mission and the long tail. Frame the podcast as a public good — a growing library of verified, accessible educational content that costs almost nothing to sustain. Include the play count, the topic diversity, the verification process. Make it easy for a program officer to say yes.
The program officer angle is important. At organizations like Mozilla or the Internet Archive, there's usually someone whose job is to give away small grants for projects that align with the mission. They want to fund things. Their performance is measured partly by how many interesting projects they support. A well-written proposal that clearly connects the podcast to their stated goals makes their job easier.
Step four: test multilingual expansion with one language — Spanish is the obvious choice — for three months. Budget about a hundred and fifty dollars per month. Use the same script pipeline, just with Spanish-language TTS and either translated or natively generated scripts. Measure the increase in total plays. If it works, expand to additional languages. If it doesn't, you've spent four hundred and fifty dollars to learn something valuable.
You've built the pipeline, which has value in itself. Once the multilingual infrastructure exists, adding a third or fourth language is much cheaper than adding the second.
Step five: do not expand to video until sponsorship revenue covers the additional costs. Video is a time sink with uncertain ROI for an audio-first show. If a sponsor specifically wants video content and is willing to fund it, that's different. But as a self-funded experiment, it's probably not the best use of limited resources.
I think there's a broader principle here: don't expand into a new medium unless you're willing to be good at it. Half-hearted video content is worse than no video content. It signals low quality to potential listeners who discover you that way. If we do video, we should do it well — which means visual demonstrations, not just a waveform over audio. And doing it well costs real time and money.
That brings us to the bigger question — not just how to fund this show, but what kind of media we're trying to build. The model we're working toward — AI-generated, verified, educational, sponsor-light — could be a template for other niche podcasts. Think about it: the barrier to creating a high-quality educational podcast has historically been very high. You need subject matter expertise, recording equipment, editing skills, distribution knowledge. AI script generation plus TTS plus automated production pipelines drops that barrier dramatically. The remaining challenges are verification — making sure the content is accurate — and distribution — getting it in front of listeners. If we solve those for this show, we've built a playbook that others can use.
The economic model matters for that playbook. If the answer is "find a wealthy patron who pays out of pocket," that's not replicable. If the answer is "a handful of mission-aligned sponsors covering very modest costs," that's a template. The Raspberry Pi Foundation funding a Spanish-language expansion of an AI-generated educational podcast — that's a story that other foundations can understand and replicate.
There's one thing we haven't addressed directly: the question of editorial independence. If we do get a sponsor or a non-profit funder, how do we ensure they never influence content?
I think the answer is to publish a sponsorship policy. Something simple and public: "Sponsors and funders have no editorial control. They do not review content before publication. They do not suggest topics. Their support is acknowledged transparently, and listeners will always know who is funding the show." Put it on the website. Make it part of the pitch. Any sponsor who balks at that is a sponsor you don't want.
That's actually a selling point for non-profits. The Internet Archive or Mozilla doesn't want to be seen as controlling content — they want to be seen as enabling independent educational media. A clear editorial independence policy protects them as much as it protects us.
The other thing we haven't mentioned: the show is already a proof of concept for something larger. Two hundred episodes, a hundred and eighty thousand plays, a growing library of verified educational content on topics from gyroscopes to diplomatic protocol — all generated for roughly the cost of a nice dinner out per month. If we can demonstrate that this model is sustainable with even modest sponsorship, the conversation shifts from "can we keep this going" to "what else could we build this way?
Which is why the multilingual expansion matters strategically, not just tactically. If the same pipeline can produce verified educational content in Spanish, Arabic, Hindi, Mandarin — suddenly we're not talking about a niche podcast. We're talking about a global educational infrastructure that costs almost nothing to operate relative to its reach.
The long tail is the secret weapon here. A traditional educational media company has to pick topics that will attract a large audience. We don't. We can make an episode about the history of the pencil, or how a sundial works, or the chemistry of soap, and it just sits there in the archive, accumulating listens over years. The marginal cost of adding another episode to the library is tiny. The marginal value, over time, is significant. That's a model that foundations should find compelling because it's sustainable in a way that grant-funded projects often aren't.
If we're wrapping this into a concrete set of recommendations — and I want to make sure we actually deliver what was asked — here's the summary. Pitch list, in priority order: Modal, Deepseek, Hugging Face, Weights and Biases, Fal or Replicate, DigitalOcean or Linode, and one of the AI education non-profits. Ad formats: "Presented by" sponsorship, "Tool of the Week" integration, and listener discount codes. No pre-roll, no mid-roll, no host-read ad copy that pretends to be excited about something. Non-profit candidates: Internet Archive, Wikimedia Foundation, Mozilla Foundation, AI Now Institute, Raspberry Pi Foundation. Expansion: test Spanish-language audio first, hold off on video until it's funded. And before any of this, implement auditable analytics.
The open question we should leave listeners with: if we do get a sponsor, how do we ensure they never influence editorial content? Should we publish a sponsorship policy? I think the answer is yes — and I think the audience would appreciate seeing it.
The model we're building — AI-generated, verified, educational, and sponsor-light — could be a template for other niche podcasts. Is this the future of independent educational media? I don't know, but I think it's worth finding out.
If you're a listener who works at one of the organizations we mentioned, or if you have a connection, reach out. We're not asking for much — just enough to keep the lights on and maybe add a few more languages.
The show notes for this episode will include a full breakdown of the costs and the pitch list. If you want to dig into the numbers yourself, they'll be there.
Now: Hilbert's daily fun fact.
Hilbert: The Nǀuu language, spoken by fewer than ten people in South Africa's Northern Cape, contains a click consonant — the bilabial click, written as a circle with a dot, like a target symbol — that linguists in the 1840s initially dismissed as a kiss sound rather than a legitimate phoneme. It took until the 1990s for the sound to be formally recognized as contrastive, meaning the difference between a word meaning "to grind" and a word meaning "to kiss" was literally a single click that Western linguists couldn't hear as language.
...right.
This has been My Weird Prompts. If you want more episodes like this one — including the spreadsheet Herman is definitely going to make after we stop recording — find us at myweirdprompts.
If you work somewhere that should be on our pitch list, you know where to find us.
