AI API costs & margins

About 32 episodes mention AI API spend, usually in the costs-and-margins segment. The most common report is that AI APIs are the largest single cost line — Diego Roshardt (about 2.5K a month), Gil Hildebrand (3.5K), CJ Zafir (2.8K), Louis Pereira, Max and Ramsri Goutham all say so — yet for text-based products margins stay high: Blake Anderson puts AI back-end cost under 3% of revenue, and GlowUp, Setter AI and Ramsri report roughly 70–90% margins. The exception is generative media: Blake warns that image generation and advanced voice push costs up, Nicolai Klemke (Neural Frames) spends about $45K a month on GPUs and video APIs — almost half of revenue — and John Rush says AI tools cut margins versus non-AI tools. Free usage is the classic trap: David Bressler ran up about $5,000 of OpenAI bills in days after going viral, and Loïc Berthelot’s free AI images scale costs with sign-ups, which is why CJ Zafir and David (Shipper) offer no free trial. Design responses include usage-capped subscriptions (Castmagic), BYOK to make a one-time price viable (Praneeth Pike) and routing tasks across models. Guests disagree on the direction of prices: one says LLM costs keep falling, Colin McIntosh says cheap credits won’t last. All figures are self-reported monthly snapshots.

Key points

• AI APIs are usually the single biggest running cost of an AI product — 8 episodes. (Diego Roshardt · Diego Roshardt (Jul 2025) ▶ 10:45) (Louis Pereira · Louis Pereira (Dec 2025) ▶ 11:53) (Max · Max (Dec 2025) ▶ 11:46) +5 more
• For text-based apps, AI costs stay small relative to revenue and margins of 70–90% are common — 8 episodes. (Blake Anderson · Blake Anderson (Dec 2024) ▶ 14:00) (Louis · Louis (Dec 2025) ▶ 10:44) (Ramsri Goutham · Ramsri Goutham (Jun 2025) ▶ 09:43) +5 more
• Free usage can blow up costs before revenue arrives: monetise early and skip free trials when inference is expensive — 5 episodes. (David Bressler · David Bressler (Feb 2024) ▶ 07:59) (CJ Zafir · CJ Zafir (May 2025) ▶ 12:39) (David · David (May 2026) ▶ 02:30) +2 more
• Generative media (image, video, advanced voice) is expensive and thins margins sharply — 4 episodes. (Blake Anderson · Blake Anderson (Dec 2024) ▶ 14:00) (Nicolai Klemke · Nicolai Klemke (Jul 2025) ▶ 09:59) (John Rush · John Rush (May 2025) ▶ 12:44) +1 more
• Route tasks across several models and tools to manage cost and quality — 4 episodes. (John Rush · John Rush (May 2025) ▶ 12:37) (Dustin Stout · Dustin Stout (Oct 2025) ▶ 08:23) (Mau Baron · Mau Baron (Jan 2026) ▶ 05:22) +1 more
• Price around the cost: usage-capped subscriptions, or BYOK to make one-time pricing sustainable — 2 episodes. (unnamed guest · Ramon Berrios & Blaine (Jun 2024) ▶ 07:01) (Praneeth Pike · Praneeth Pike (Oct 2025) ▶ 07:08)

Where guests agree

• Track AI spend against price continuously, because costs scale with usage (unnamed guest · Ramon Berrios & Blaine (Jun 2024) ▶ 07:01) (Louis · Louis (Dec 2025) ▶ 10:44) (David Bressler · David Bressler (Feb 2024) ▶ 07:55) +1 more

Where guests disagree

Will AI API costs keep falling? ⚠️

• Yes — LLM costs are constantly going down (unnamed guest · Eyal & Yahli (Nov 2025) ▶ 09:06)
• No — credits are cheap now because providers are buying users; the window may close, so move fast (Colin McIntosh · Colin McIntosh (Jan 2025) ▶ 02:46)

Frameworks

• Zero-marginal-cost one-time-purchase AI app — How Praneeth restructured the app so a one-time price is sustainable. (Praneeth Pike · Praneeth Pike (Oct 2025) ▶ 04:03)
  • Step 1: bring your own keys - users pay AI usage
  • Step 2: no server costs - run on users’ computers via a desktop app
  • Sell once as a lifetime/one-time purchase
  • Raise price with user milestones ($29 under 100 users, $49 after)

Tactics

• Skip free trials when inference is expensive — No trials because top AI models are not cheap; sell monthly and discounted yearly memberships. (CJ Zafir · CJ Zafir (May 2025) ▶ 12:39)
• Add a donation link to recoup costs before monetising — A Stripe link saying he built the app, it went viral, feel free to donate; he made a few thousand back. (David Bressler · David Bressler (Feb 2024) ▶ 08:12)
• Watch AI costs against price — Continuously track spend on LLMs and transcription to keep margin on a usage-capped subscription. (unnamed guest · Ramon Berrios & Blaine (Jun 2024) ▶ 07:01)
• Offer several LLM providers inside your product — Because the best model changes monthly, he wires in three model APIs and lets users choose. (John Rush · John Rush (May 2025) ▶ 12:37)
• Use different AI tools for different tasks to spread token costs — Cursor as daily driver, Codex CLI for complex tasks, ChatGPT for questions and easier problems. (Mau Baron · Mau Baron (Jan 2026) ▶ 05:22)
• Bundle many AI subscriptions into one — Aggregate top LLMs via OpenRouter and image/video models via a similar service behind one subscription. (Dustin Stout · Dustin Stout (Oct 2025) ▶ 08:23)

Stories & examples

• $5,000 OpenAI bill after going viral: After the Reddit virality he checked OpenAI billing and found he had spent about $5,000 in API costs in days; donations and sponsorships gave him the motivation not to shut it down. Lesson: Viral free usage of an AI app can create large costs before revenue; monetise early. (David Bressler · David Bressler (Feb 2024) ▶ 07:59)
• Bible Buddy on WhatsApp: Launched as an emotional AI therapist for Christians on WhatsApp via a third party because he lacked technical skills; WhatsApp’s per-conversation fees pushed costs to about 15k a month, so he moved to his own infrastructure on iOS and Android with an Upwork engineer, then sold it to a buyer met through the My First Million Facebook community. Lesson: Check platform per-use costs before launching on them. (Lotts Ezeike · Lotts Ezeike (Nov 2025) ▶ 03:08)
• To sell a one-time purchase without ongoing costs, he made it bring-your-own-keys so users pay for AI usage, and built a desktop app so users’ computers carry server costs. Many more people paid, gave serious feedback and believed in the product, leading to about $80,000 in six months. (Praneeth Pike · Praneeth Pike (Oct 2025) ▶ 04:03)
• About $45,000 a month on GPUs and text-to-video APIs (almost half of revenue), about 5K on servers and hosting, about a thousand a month on Cursor, plus smaller tools; he says there isn’t much profit in the end. (Nicolai Klemke · Nicolai Klemke (Jul 2025) ▶ 09:59)

Numbers

• GPU and text-to-video API spend: 45000 USD/month / month (self-reported, approx.) (Nicolai Klemke · Nicolai Klemke (Jul 2025) ▶ 09:59)
• OpenAI API costs run up in days: 5000 dollars (self-reported, approx.) (David Bressler · David Bressler (Feb 2024) ▶ 07:55)
• SEObot margin, monthly revenue and monthly API cost: 70% margin; 100K revenue; 30K APIs percent; currency unstated per month / month (self-reported, approx., currency unstated) (John Rush · John Rush (May 2025) ▶ 12:44)
• Monthly costs: Cursor / Vercel / AI APIs: 20 / 40 / 2,500 currency unstated / month (self-reported, approx., currency unstated) (Diego Roshardt · Diego Roshardt (Jul 2025) ▶ 10:42)
• AI compute cost: 3500 USD/month (self-reported, approx.) (Gil Hildebrand · Gil Hildebrand (Dec 2025) ▶ 13:50)
• Monthly costs: back end and AI: ~1,000 back end; ~500 AI per month (AI in ‘bucks’) (self-reported, approx.) (Louis · Louis (Dec 2025) ▶ 10:55)
• Cost to create a full book with GPT-3: 3 USD (self-reported, approx.) (Joe Popelas · Joe Popelas (Aug 2023) ▶ 06:35)
• Transcription cost: about 1 per hour; 2,000-3,000 per month USD (self-reported, approx.) (Jure Sotosek · Jure Sotosek (Apr 2025) ▶ 14:18)

Caveats & warnings

• Margins are thin: GPU/API costs are almost half of revenue and ‘not much profit in the end’. (Nicolai Klemke · Nicolai Klemke (Jul 2025) ▶ 10:25)
• AI costs rise above the low single digits if you do image generation or advanced voice mode. (Blake Anderson · Blake Anderson (Dec 2024) ▶ 14:00)
• Costs and margins depend on user usage. (Louis · Louis (Dec 2025) ▶ 10:44)
• The cheap-AI window may close; move quickly. (Colin McIntosh · Colin McIntosh (Jan 2025) ▶ 02:37)

How much to trust this

• Evidence: 32 episodes, 33 independent speakers (repeat guests counted once; Pat and Gus counted as one voice).

Blueprints that use this

• AI product with a moat