AI Has a Subsidization Problem - CourtneyR.dev's 2nd Brain

AI Has a Subsidization Problem ![rw-book-cover](https://i.ytimg.com/vi/w62xTVuyu3s/sddefault.jpg) 🔗 **Source:** [AI Has a Subsidization Problem](https://www.youtube.com/watch?v=w62xTVuyu3s) > [!note]+ Document Notes > tags: #ai #subsidization #business-models #compute-costs #developer-tools > author: Theo - t3dotgg > source: YouTube > reference: https://www.youtube.com/watch?v=w62xTVuyu3s > date created: 2024-12-28 12:00 > > # 🏷️ Title: AI has a subsidization problem >  ## 🧰 Action Inbox - [H1] Evaluate current AI tool subscriptions for actual value vs subsidized pricing — 00:00-05:00 #business-strategy - [H2] Prepare for end of free inference tiers across AI providers — 15:30-20:00 #cost-planning - [H3] Consider switching to direct API usage for better control — 25:45-28:00 #technical-infrastructure ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [Gemini CLI Changes](https://youtu.be/w62xTVuyu3s) - [GitHub Copilot Student Plan Updates](https://youtu.be/w62xTVuyu3s) ## 📝 Summary - AI companies are ending subsidized inference programs, starting with Google's Gemini - Free and heavily subsidized AI services were economically unsustainable business strategies - Companies used subsidization to steal customers but attracted non-paying users instead - Token usage has increased 10x+ due to reasoning tokens and tool calls, counteracting price reductions - The era of "free compute" is ending as companies realize ads can't cover inference costs ## 📖 Main Content The AI industry is experiencing a fundamental shift as companies end their subsidization programs. Google has become the first major player to significantly restrict free access to their Gemini models, cutting off free users from Pro models entirely and limiting paid subscribers' quotas. This represents the beginning of the end for heavily subsidized AI inference. The core issue is economic reality: there is no such thing as free compute. While per-token costs have decreased, actual usage costs have skyrocketed due to reasoning tokens, tool calls, and longer generations. A single message that once generated 200 tokens might now generate 200,000 tokens, requiring significantly more GPU time and resources. The subsidization strategy was primarily designed to steal customers from competitors, not create new markets. However, this approach backfired by attracting users who couldn't afford to pay, rather than converting valuable customers who would eventually upgrade to paid tiers. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/w62xTVuyu3s?t=0) — [H1] Google ends Gemini subsidization - Key idea: Google restricts free tier access to Pro models and limits paid quotas - Evidence: "Important update to GitHub C-Pilot for students... some premium models including GPT 5.4 for and Claude Opus and Sonnet will no longer be available" #### ⏱️ [03:45](https://youtu.be/w62xTVuyu3s?t=225) — [H2] No such thing as free compute - Key idea: All inference has real costs that must be covered somehow - Evidence: "Just running local models on my own GPUs in my apartment costs me hundreds of dollars a month in electricity alone" #### ⏱️ [06:30](https://youtu.be/w62xTVuyu3s?t=390) — [H2] Token usage explosion negates price drops - Key idea: While per-token costs decreased 10x, usage increased even more - Evidence: "If one chat message used to generate 200 tokens and now it generates up to 200,000 tokens that's also more time being spent on the GPU" #### ⏱️ [09:15](https://youtu.be/w62xTVuyu3s?t=555) — [H3] Message cost variability is extreme - Key idea: Individual messages can vary by 400x in token usage and cost - Evidence: "Some messages cost literally 0.001. Other messages can cost multiple dollars" #### ⏱️ [12:00](https://youtu.be/w62xTVuyu3s?t=720) — [H2] Google's desperate customer acquisition strategy - Key idea: Google subsidized heavily to steal OpenAI customers but failed - Evidence: "I don't think I've ever met a person who shut down their $20 a month Chashibbt subscription to go subscribe to $20 a month of Gemini" #### ⏱️ [15:30](https://youtu.be/w62xTVuyu3s?t=930) — [H1] Free users aren't valuable customers - Key idea: Users who can only afford free tiers will never become paying customers - Evidence: "Those customers aren't valuable enough to get free inference. There is no world in which that continues" #### ⏱️ [19:45](https://youtu.be/w62xTVuyu3s?t=1185) — [H3] Ads cannot cover inference costs - Key idea: Advertising revenue is insufficient to subsidize expensive AI inference - Evidence: "20,000 days of my content being watched is worth about $9,000 with ads... not even enough to pay my editor" #### ⏱️ [25:45](https://youtu.be/w62xTVuyu3s?t=1545) — [H2] Data collection has limited value - Key idea: Training data from free users isn't valuable enough to justify costs - Evidence: "If you can't afford a $20 a month sub your inference is worth less than the average person" #### ⏱️ [30:20](https://youtu.be/w62xTVuyu3s?t=1820) — [H1] Customer theft is the real subsidization goal - Key idea: Companies subsidize to steal competitors' customers, not grow markets - Evidence: "There's two reasons somebody will change... something else that's better or something else that's good enough and cheaper" #### ⏱️ [35:10](https://youtu.be/w62xTVuyu3s?t=2110) — [H2] Death spiral of free-only users - Key idea: Products with only free users who can't afford alternatives are doomed - Evidence: "If your only users are people who are using your thing because they can't afford to use the better thing, you're fucked" #### ⏱️ [40:30](https://youtu.be/w62xTVuyu3s?t=2430) — [H3] Enterprise subsidization model works differently - Key idea: High-value users can subsidize others when properly structured - Evidence: "That $200 a month plan can be used for up to $5,000 of compute per month" #### ⏱️ [45:15](https://youtu.be/w62xTVuyu3s?t=2715) — [H2] Anthropic's strategic lock-in approach - Key idea: Anthropic subsidizes to create vendor lock-in, not choice - Evidence: "They want to get people off of cursor where every model can be used and onto claude code where only enthropic models can be used" #### ⏱️ [50:00](https://youtu.be/w62xTVuyu3s?t=3000) — [H1] Google's incompetent approach - Key idea: Google's subsidization was accidental mismanagement, not strategy - Evidence: "Google subsidized so hard they fucked over their own developers" ## 📚 References - [Gemini CLI Official Updates](https://youtu.be/w62xTVuyu3s) - [GitHub Copilot Student Plan Changes](https://youtu.be/w62xTVuyu3s) - [T3 Chat Platform](https://youtu.be/w62xTVuyu3s) ## 🗣️ Paraphrase - AI companies are ending free inference programs because they're economically unsustainable and attract non-paying users - Token usage has exploded due to reasoning and tool calls, making individual requests much more expensive despite lower per-token costs - Subsidization was primarily a customer acquisition strategy to steal market share, but most companies failed to convert free users to paying customers ## 📰 Newsletter The era of free AI inference is rapidly coming to an end, and developers need to prepare for this shift. Google's recent restrictions on Gemini access signal that other providers will likely follow suit as the economics of subsidized AI services prove unsustainable. For WordPress developers using AI tools for code generation, content creation, or site optimization, now is the time to evaluate which paid subscriptions provide genuine value. The $200/month enterprise tiers from providers like Anthropic and OpenAI offer significant compute allowances that can support serious development work. Consider establishing direct API relationships rather than relying on intermediary services that may lose access to premium models. This ensures more predictable costs and feature availability for your WordPress projects. Resources: [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/), [WP-CLI](https://wp-cli.org/) ## ✅ Action Items #### DevRel - Developer advocacy: Create cost-aware AI tooling guides, demonstrate ROI calculations for paid tiers - Education: Workshop on transitioning from free to paid AI services, budget planning sessions - Developer experience: Document migration paths from subsidized to direct API usage - Developer marketing: Position paid AI tools as professional development investments - Community management: Survey community AI tool usage, gather feedback on pricing sensitivity #### OSPO - Policy & Governance: Establish AI service procurement policies, vendor evaluation criteria - License Compliance: Review AI service terms for open source project usage restrictions - Software Inventory: Catalog AI dependencies, assess vendor lock-in risks - Developer Support: Training on cost-effective AI tool usage, budget management - Strategic Alignment: Align AI tool investments with organizational objectives #### Community - Maintainers: Provide guidance on sustainable AI tooling for project development - Contributors: Create pathways for AI-assisted contributions without cost barriers - Sponsors: Consider funding AI tool access for key contributors - Ecosystem: Coordinate bulk purchasing or enterprise agreements across projects #### WordPress - Core: Evaluate AI tool usage for core development, establish sustainable tooling budget - Documentation: Update developer resources with current AI tool landscape and costs - Community: Provide guidance for meetups and WordCamps on AI tool selection - Training: Include AI cost management in developer education curriculum ## 🧪 Hypotheses - AI companies will consolidate around 2-3 major providers within 18 months as subsidization ends - Enterprise-focused AI tools will see higher retention rates than consumer-focused free tiers - Direct API usage will become more cost-effective than wrapper services for heavy users - Smaller AI startups will struggle to compete without subsidization strategies ### 📊 Methodology - Track pricing changes and feature restrictions across major AI providers over 6 months - Survey developer communities on AI tool usage patterns and budget allocation - Analyze correlation between AI tool costs and user retention rates - Compare total cost of ownership for direct API vs. wrapper services ### ❓ Questions - Which AI providers will maintain sustainable free tiers beyond 2024? - How will the end of subsidization affect AI adoption in smaller development teams? - What new business models will emerge to replace the subsidization approach? ### 🔮 Future Research - Economic sustainability models for AI inference services - Impact of inference cost increases on developer productivity - Market consolidation patterns in the AI tooling space ### 💡 Implications - Practice: Developers need to budget for AI tools as essential infrastructure costs - Policy: Organizations should establish AI procurement and budget policies - Security: Reduced reliance on free tiers may improve data security and privacy - Education: Training programs need to include AI cost management and tool selection - Funding: Open source projects may need dedicated AI tooling budget allocations ### 🧠 Learning Styles Assessment Reading level estimate: 10th Grade (Flesch-Kincaid) #### 👀 Visual learners - Present: Cost comparison charts, user segment diagrams - Improve: Timeline of subsidization changes, economic flow diagrams #### 🔊 Aural learners - Present: Video format with clear narration and examples - Improve: Podcast version focusing on key economic concepts #### 🗣️ Verbal learners - Present: Clear definitions of subsidization, inference costs, business models - Improve: Analogies comparing AI costs to other utility services #### 🧩 Kinesthetic learners - Present: Real examples with actual cost calculations and usage scenarios - Improve: Interactive cost calculators, hands-on budget planning exercises #### 🧠 Logical learners - Present: Clear cause-and-effect reasoning, economic analysis frameworks - Improve: Decision trees for AI tool selection, ROI calculation templates #### 🧑‍🤝‍🧑 Social learners - Present: Community examples, shared experiences from developers - Improve: Discussion forums for cost-sharing strategies, group buying coordination #### 🧘‍♀️ Solitary learners - Present: Individual analysis and reflection on personal AI tool usage - Improve: Self-assessment worksheets for AI tool needs and budget planning #### 🌿 Naturalistic learners - Present: Market ecosystem analysis, natural evolution of business models - Improve: Connections to broader technology adoption lifecycle patterns ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear headings and logical content structure - Improve: More descriptive alt text for any charts or diagrams #### Keyboard navigation - Present: Standard video controls and text navigation - Improve: Timestamp navigation shortcuts #### Contrast and readability - Present: Standard text contrast in video player - Improve: High contrast mode support for accompanying materials #### Media and captions - Present: Video format with transcript provided - Improve: Time-synced captions, downloadable transcript with timestamps ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-centric perspective on AI costs and business models - Improve: Global perspective on AI accessibility and cost barriers #### Inclusive language - Present: Technical jargon well-explained, casual but professional tone - Improve: More sensitivity around economic privilege and access barriers #### Neurodiversity support - Present: Clear structure with timestamps and sections - Improve: Content warnings for strong language, summary sections #### Representation - Present: Male perspective from established developer - Improve: Include perspectives from developers with different economic situations ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear economic logic about unsustainable subsidization models - Inductive strength: Good use of specific examples (Google, Anthropic) to support broader claims - Causal reasoning: Well-established links between token costs, usage patterns, and business viability - Analogical strength: Effective comparison to other utility and subscription services #### Fallacies - Ad hominem: Some dismissive language toward free users could be seen as attacking rather than analyzing - False dichotomy: Somewhat oversimplifies the range of sustainable business models #### Tone of voice - Overall tone: Authoritative and somewhat combative, with strong opinions backed by economic reasoning ### 🛡️ DARVO Analysis - Deny: No significant denial patterns present - Attack: Some attacks on "poor users" and free tier users, though framed as economic analysis - Reverse victim/offender: No reversal patterns identified - Emotional cues and rhetoric: Strong language used to emphasize economic realities, some dismissive framing of certain user segments ## 🎯 Conclusion The AI subsidization era is ending as companies realize that free inference models are economically unsustainable. Developers and organizations must prepare for a transition to paid tiers and direct API relationships. While this represents increased costs, it also signals market maturation and more sustainable business models that will ultimately benefit serious AI tool users through better service reliability and feature development. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] AI subsidization models are economically unsustainable > [!quote] [AI has a subsidization problem](https://youtu.be/w62xTVuyu3s) by [[Theo - t3dotgg]] via [[YouTube]]. Published 2024. Accessed 2 --- ## Full Transcript [00:00] We want to be transparent that we're [00:01] making changes to Gemini CLI that may [00:03] impact your workflow, including more [00:04] robust detection of policy violating use [00:06] cases, prioritizing traffic for certain [00:08] types of accounts and restricting Gemini [00:11] Pro models for free tier users. [00:12] Important update to GitHub C-Pilot for [00:14] students. Your complimentary co-pilot [00:16] access will be managed under a new [00:18] GitHub copilot student plan. As part of [00:20] this transition, however, some premium [00:22] models including GPT 5.4 for and Claude [00:24] Opus and Sonnet will no longer be [00:26] available for self- selection under the [00:28] GitHub copilot student plan. Open code [00:30] 1.3.0 will no longer load the Claude Max [00:34] plugin. We did our best to convince [00:36] Anthropic to support developer choice, [00:38] but they sent lawyers. Hi Google team. [00:40] I'm a Google AI pro subscriber and I [00:42] believe there's a bug or undocumented [00:43] change in how quotas are being applied [00:45] to my account. We should receive high [00:47] generous quotas refreshed every 5 hours [00:49] with higher weekly rate limits than free [00:51] users. However, my current anti-gravity [00:53] UI showing much longer cool downs. It [00:56] looks like the subsidization era might [00:57] be coming to an end faster than we [00:59] expected. I'm pretty surprised if I'm [01:02] being honest. I did not think Google [01:03] would be the first company to really tap [01:05] out here and start cutting off the free [01:07] and subsidized access to users that are [01:10] still hyped on Gemini stuff for some [01:11] reason. It's kind of crazy to see how [01:14] quickly the tides have turned here. from [01:16] giving out thousands of dollars of [01:17] inference for hundreds of dollars to [01:19] locking out free users entirely and [01:21] massively restricting what student plan [01:23] users can have on things like GitHub [01:24] Copilot. This is the start of the end. [01:27] But in order to understand and [01:28] appreciate the end, we probably need to [01:30] better understand the start, too. How we [01:32] got here, what this all means, and what [01:34] the future of the economics of AI [01:36] development stuff is. Are these [01:38] companies even going to exist in a few [01:39] years? I have no idea. Thankfully, none [01:42] of them are paying me for any of my [01:44] coverage here. So, I'm going to do my [01:45] best to cover this all in an unbiased [01:47] way and help you better understand where [01:49] the tools and the subscriptions we're [01:50] paying for now are going. But in order [01:52] to pay for them, I need some money. So, [01:53] we're going to do a quick break for [01:54] today's sponsor. Sometimes the popular [01:56] solution isn't very good. And I would [01:58] confidently say that's the case with [01:59] GitHub actions. The fact that we run [02:00] most of our CI through them is pathetic. [02:02] Even if you're using something awesome [02:04] like today's sponsor, Dep Depot, they [02:06] make your actions way faster and also [02:08] your Docker builds. But they are kind of [02:09] hampered by what GitHub allows because [02:11] GitHub actions to be frank suck. If only [02:14] somebody who knew how to make actions [02:16] better was to take the time to make [02:17] better action. Oh, what's that on my [02:19] screen? Now available. Depotci. Oh, [02:23] depot fixed GitHub actions by killing [02:25] GitHub actions. I really like how they [02:27] framed this year. They did everything [02:28] they could to make GitHub actions as [02:30] good as possible, but there was always a [02:32] ceiling they had to hit. And now they've [02:34] hit it. Turns out they can only improve [02:36] about 30% of the system. But the other [02:38] 70% was outside of their control. That [02:40] makes it slower. That makes it harder to [02:42] interface with. That makes it impossible [02:43] for your agents to do anything. And I'm [02:45] sure we've all had to copy paste errors [02:47] from our CI back over to our agents. [02:49] What if there was a CLI they could use [02:50] to actually get that information? What [02:52] if there was more reliable ways to SSH [02:54] into the boxes that are running your [02:56] stuff so you can actually figure out [02:57] what's going wrong? What if there were [02:58] dashboards that gave you actually useful [03:00] information and skills you could hand [03:02] your agents to take advantage of all of [03:04] this? What if GitHub actions were [03:05] reinvented today by people who knew what [03:07] they were doing and wanted to enable [03:08] agents to be more successful at building [03:10] in your real code bases? And what if you [03:12] can migrate everything over with two [03:14] commands? Yes, it's actually that easy [03:16] to get started. It's already way faster [03:18] and better to use, not just for yourself [03:20] as a developer, but for your agents, [03:21] too. Stop fighting GitHub and get back [03:23] to work at soyv.link/depo. [03:25] There's no such thing as free compute. [03:28] And I want to be very, very strict about [03:31] this fact. People seem to think that AI [03:33] is just free once the models are [03:35] created. And they're not. Just running [03:38] local models on my own GPUs in my [03:40] apartment costs me hundreds of dollars a [03:42] month in electricity alone. These things [03:44] do cost money. And if you're giving them [03:47] out for free, there has to be a reason [03:49] for that. It could be that you're making [03:50] money on ads. It could be that you're [03:52] trying to sway a customer with a trial. [03:54] It could be that you're collecting a ton [03:56] of data. And chances are it's some [03:57] combination of all of those things in [03:59] order to make this economically viable, [04:02] but we're not getting to a place where [04:03] giving out free inference in general [04:06] makes a lot of sense. And this is a [04:07] theory we've had for a while. The price [04:09] of Frontier Intelligence has been going [04:11] down somewhat consistently. This is a [04:13] chart that shows how the cost of [04:15] different models dropped from 2023 [04:17] onwards and things like 40 and 40 mini [04:20] were upwards of 10 times cheaper than [04:22] the equivalent of GPT4 32K back in early [04:26] 2023. And it seemed like this was going [04:28] to stay the trend especially with [04:29] openweight models coming out and being [04:31] even cheaper. But there was something we [04:32] didn't expect. We didn't expect the [04:34] amount of inference we are doing to go [04:36] up exponentially. If you were to do a [04:38] similar chart here for the number of [04:40] tokens a model generated in order to [04:42] respond to a question in 2023 versus [04:45] now, the amount of tokens being used has [04:47] gone up at least 10x because now we're [04:50] prompting in things like our code bases. [04:52] And the models aren't just oneshotting a [04:54] response, they're doing tool calls, [04:56] which means more additional responses, [04:58] more additional steps being done to go [05:00] collect data and then another chat [05:02] effectively happening for each step [05:04] along the way. reasoning tokens [05:06] massively increasing costs here as well [05:08] massively increasing the amount of [05:09] tokens being generated the amount of [05:11] money that we are spending not just on [05:13] AI in general because we're using AI [05:14] more but for a given prompt has more [05:17] than counteracted the amount the costs [05:19] have gone down but more importantly the [05:21] amount we need the GPUs to be doing [05:23] things for has gone up if one chat [05:26] message used to generate 200 tokens and [05:28] now it generates up to 200,000 tokens [05:30] that's also more time being spent on the [05:32] GPU and if you're using the GPU than [05:34] somebody else isn't and that costs money [05:36] inherently. So while intelligence is [05:39] getting cheaper in the sense of per [05:41] token sometimes in some cases it's not [05:43] enough and the demand for doing these [05:45] long generations that take more GPUs and [05:48] more time on the GPUs is consistently [05:50] going up. So where does Gemini come in [05:52] here? Well, it comes in because of the [05:54] subsidization wars. When AI models first [05:57] kind of became a thing, we were trying [05:59] to figure out how to charge for them [06:01] because users at the time and even now [06:03] to an extent don't really understand [06:05] tokens and the idea being build per sets [06:08] of characters didn't make sense to them. [06:11] So many products including ourselves [06:12] with T3 chat decided to bill based on [06:14] the number of messages. That's a [06:17] problem. If I go to a service like T3 [06:19] Chat and I ask it, what's 2 plus 2? you [06:22] get an answer generated very quickly [06:24] with not too many tokens, only 11 tokens [06:26] to answer this. But if I ask it to write [06:30] 15 poems about React, I'm not going to [06:33] do 15. I'll do five. Be conservative [06:35] here. This message was 434 tokens [06:38] generated. That's 40x more tokens for a [06:41] roughly similar question length because [06:43] the thing I'm asking for is just more [06:45] complex here. And the answer is going to [06:47] be more. And if you think I kind of [06:50] exaggerated this gap here with the 40x [06:52] example, I didn't. In real world [06:54] experience, I could say that the gap [06:56] between the bottom 10% most token used [06:59] message and the top 10% most tokens used [07:01] in a message is closer to a 400x gap. So [07:04] some messages cost literally 0.001. [07:08] Other messages can cost multiple [07:10] dollars. And yes, you can send a message [07:13] on T3 Chat for eight bucks a month where [07:15] the individual message cost us on the [07:17] API a dollar, an eighth of your sub with [07:21] one prompt. It's kind of insane, but [07:23] that's just how the models work and how [07:25] these things naturally can cost more [07:28] when more is happening. At the same [07:30] time, these business customers are [07:32] growing their needs. The desire for [07:34] these companies to compete with each [07:35] other is going up too as they fight for [07:37] end users. Open has pretty much won the [07:39] end user right now with the $20 a month [07:41] chasht subscription that makes up 70% of [07:43] their revenue. They're killing it there. [07:45] I don't think I've ever met a person who [07:47] shut down their $20 a month Chashibbt [07:49] subscription to go subscribe to $20 a [07:51] month of Gemini. No one does that. The [07:53] closest I've seen is people who use the [07:56] free tier of Gemini and realize they [07:58] don't really need Chashibbt anymore. So, [07:59] they go over and cancel it. And I get [08:01] that. I understand that. And this is [08:03] kind of the bet Google wants to make is [08:06] can they make intelligence good enough [08:07] and cheap enough that every user will [08:10] just start using Google's intelligence [08:12] instead eventually killing all the [08:13] competition and then Google owns the [08:15] whole AI world and they can charge [08:16] whatever they want for the companies [08:18] doing the categorization. The problem is [08:20] that we're all fighting over the same [08:21] GPUs. Every one of these companies has [08:23] an amount of infrastructure they can use [08:25] to serve requests for their business [08:27] customers, for their research labs, [08:28] generating things with AI and making [08:30] their own models and making their own [08:32] whatever they're doing and the end users [08:34] who are paying the whatever a month to [08:36] go use the models in products like [08:38] anti-gravity or T3 code or just using it [08:41] in their chat interface directly. And [08:43] since Google in particular desperately [08:45] wants to steal customers off of OpenAI [08:48] and off of Anthropic, they've been way [08:50] too generous with subsidizing usage in [08:53] their products. I cannot tell you how [08:55] many times I have seen comments from [08:58] lower income people in my content, be it [09:00] Twitter, Twitch, YouTube, or wherever, [09:03] saying that I need to stop being so mean [09:05] to anti-gravity because it's the best [09:07] option for poor people who can't afford [09:09] a subscription for something better. to [09:11] which I would usually bite my tongue [09:12] because I know the reality here. Those [09:14] customers aren't valuable enough to get [09:16] free inference. There is no world in [09:18] which that continues. And we are now at [09:20] the end of that. I had a feeling it [09:22] would come, but I honestly thought [09:24] Google's billions upon billions of [09:26] dollars would sustain this a bit longer. [09:28] Google got users through giving out free [09:31] inference, but that meant the majority [09:32] of their users were using free [09:34] inference, which meant that they were [09:36] eating a ton of costs and had a lot of [09:38] GPUs that were being used by users who [09:40] would never spend actual money when they [09:42] could be used by people who are [09:44] currently spending actual money. And [09:45] this has affected us. We've had times [09:47] where people are using paid things on T3 [09:50] Chat where we're paying API prices and [09:51] they send a message and 3.1 just fails [09:54] to respond because it's overloaded. or [09:56] even I on the $250 a month Google plan [09:59] that I subscribed to forever ago for [10:01] content and probably should unsubscribe, [10:03] but it's fun to have it there to [ __ ] [10:05] about things like this. I couldn't use [10:07] the new Gemini 3.1 Pro model when it [10:09] dropped in the $250 a month subscription [10:11] and the Gemini CLI. And when asked, they [10:14] cited this as a capacity issue because [10:16] they're giving it away for free too [10:17] much. There is no such thing as free [10:19] inference. Either you're the product and [10:20] you won't be for long or something [10:22] illegal is happening. So, if you were [10:24] picking these tools because they're free [10:26] and you can't afford it and you didn't [10:28] use the two months of thousands of [10:30] dollars of free inference, mind you, [10:31] like you're getting thousands of dollars [10:33] of compute for free during that window. [10:35] If that wasn't enough for you to find a [10:37] way to make 20 bucks a month, that's [10:38] your problem, not Google's. Seriously, [10:40] like this, this was just going to [10:42] happen. I am surprised it did so quick. [10:45] But now there is zero reason to be using [10:48] anti-gravity or the Gemini CLI because [10:50] they are not subsidizing anywhere near [10:52] as hard. They are kicking free users off [10:55] of the pro models entirely. And as much [10:57] as I am a Gemini Flash simp to an [10:59] extent, they are not good enough for [11:01] day-to-day code, the idea of using [11:03] something like Gemini 3 Flash on the [11:05] free tier or Gemini CLI for doing real [11:07] code makes me feel somewhat sick. It's [11:09] like using a toddler's like toy keyboard [11:12] to try and build a real application or [11:14] like using a FisherPrice karaoke machine [11:16] to record an album. Like it's just it's [11:18] not there. It feels silly. The craziest [11:20] line here though is around direct [11:23] control. For direct control over your [11:25] own quotas and billing, use your own [11:27] paid API key from AI Studio or [11:29] Vert.Ex.ai. This is them saying directly [11:32] on the official Gemini CLI account, [11:34] we're going to cut down your usage on [11:36] the paid subscriptions. So if you want [11:39] more usage, just go use an API key. [11:41] That's crazy. That's insane. They're [11:44] cutting you guys off hard. So how do we [11:46] get here? Why would these companies all [11:48] do free and subsidized inference? It [11:50] seems to just bring drama. Like remember [11:52] the drama in July of last year when [11:55] cursor switched away from a per message [11:57] to a billing on usage model and everyone [12:00] freaked the [ __ ] out. It seems like [12:01] these things just cause problems. Well, [12:04] the first reason you would do this is [12:05] you can make money elsewhere. probably [12:07] ads. If you have the ability to make [12:09] enough money on ads to cover the [12:11] inference being done, awesome. But I [12:14] think people massively overestimate how [12:16] much money you can make on ads. One of [12:18] my favorite ways to demonstrate how [12:19] little money there is in ads is myself. [12:22] My YouTube channel is pretty popular. We [12:24] get around 3 million views per month [12:26] from experienced developers. So, our CPM [12:28] is relatively high. In the last 28 days, [12:32] my channel got 465,500 [12:34] hours watched. So with 672 hours in the [12:38] 28 days, that puts us at 693 [12:42] hours per hour. So every time an hour [12:45] passes, 693 hours of my content are [12:49] watched. For those who have been [12:50] subscribed and watching for a long time, [12:51] we appreciate you immensely. These [12:53] numbers are insane. They're still hard [12:54] for me to just process. that we're [12:57] nearly at a thousand hours watched per [12:59] hour of time that passes is just [13:02] mind-boggling. That at any given moment, [13:04] on average, 600 to 700 people are [13:06] watching Theo talk is crazy. So, how [13:09] much money do you think I make off of [13:10] that per month? Remember, experienced [13:13] developers, high CPMs, because these [13:15] developers are experienced. How much [13:17] money do you think I get off of YouTube [13:19] ads per month? And again, do not answer [13:21] chat if you already know the answer. Get [13:22] a number in your head. I'm going to hop [13:24] over to Twitch chat so we see what [13:26] people here think. Okay, we're seeing [13:27] 40,000, 20,000, 10K to 20K, 500 bucks [13:31] from short used to work at Twitch, [13:32] you're cheating. The answer, it's about [13:35] $9,000 a month. It varies between 5 to 9 [13:38] depending on a lot of different things. [13:40] I want to emphasize the absurdity here. [13:42] This is 20,000 days of my content [13:45] watched in 28 days. So 20,000 days of my [13:49] content being watched is worth about [13:51] $9,000 with ads. That's not even enough [13:54] to pay my editor. Ads on YouTube are not [13:57] a real way to make money even at these [14:00] types of largecale operations. And this [14:03] is revenue that is being split with me. [14:05] So Google's complete ad rev here is [14:07] probably double this. It's believed that [14:09] it's a 50-50 split. There's no way for [14:11] us to truly know for a fact that's the [14:13] case, but I I take them at their word. [14:15] It probably is. So Google off of my [14:17] content theoretically made upwards of [14:20] $18,000 [14:22] total in that time. So if you think ads [14:25] can cover the costs of inference, you're [14:28] very very dumb. Like that's all I can [14:30] say there. If we reframe this to be [14:33] dollars per view, we are at.28 [14:37] per view. Do you understand how small [14:40] that is? And remember, an individual [14:42] message can cost a dollar or more. So, [14:45] if you think a little ad that makes a [14:46] fraction of a cent underneath your [14:48] prompt is enough to justify the prompt [14:50] being free, I have a whole bunch of [14:51] bridges down the street I would love for [14:52] you to take a look at. Seriously though, [14:54] like there is no money in advertising, [14:57] at least on an individual level. People [14:59] see companies like Google pulling in [15:01] billions for ads because they're doing [15:03] it at the scale of billions. To make a [15:05] billion dollars on ads, first you must [15:07] send a trillion ads out. And if each of [15:09] those instances could cost a dollar, [15:12] you're just going to lose money. So [15:13] there's no world in which.3 cent ads are [15:17] going to cover multiple dollars of [15:19] prompts. The world just does not work [15:21] that way. So if you think ads are enough [15:23] of a way to fully subsidize the cost of [15:26] these things that you should be able to [15:27] get free inference because advertising, [15:29] you do not know how any of this works. I [15:31] just want to make sure that the scale of [15:32] this is clear because people seem to [15:34] think ads make a lot more money than [15:36] they do. This is also why we're so [15:37] thankful to have awesome sponsors on [15:39] this channel because they make this [15:41] work. If it wasn't for sponsors, I would [15:43] not be able to do this because revenue [15:44] we get from people paying me on Twitch [15:46] and from ads on YouTube is just barely [15:48] enough to cover one of my team members. [15:51] Far from the entire team that we have to [15:52] run the channel with. It's certainly far [15:54] from me making a scent on it. In fact, [15:56] I'm so thankful for my sponsors, I'm [15:57] going to tell you about one right now. [15:59] AI's gotten really good at building [16:00] apps. Well, web apps at least. Once it [16:02] comes to mobile, things get a lot [16:04] messier. Even just trying to release the [16:06] apps can be obnoxious. You want to take [16:08] advantage of the cool things your phone [16:10] can do. be it the AR side or the camera [16:12] or all the other magical stuff embedded [16:14] in the Swift bindings for your iPhone, [16:16] you're kind of out of luck. Well, you [16:18] were before Ror Max dropped because this [16:21] is crazy. I'm actually like genuinely [16:23] blown away. I've been setting it up to [16:25] play with for this ad and I think I'm [16:27] addicted. It's hilariously easy to make [16:30] real applications with this. The one I [16:32] made is a bit goofy and you might notice [16:33] that you can't really see it here. [16:35] That's because I have it plugged into my [16:37] phone. Ror now ships a Mac app so that [16:39] you can link your phone and your [16:41] computer together as you use cloud code [16:43] in the background by the way in order to [16:45] build real native iPhone applications. [16:48] This is a real Swift app that was built [16:50] and it wasn't even built on my machine. [16:51] It was built in the cloud and then they [16:53] use their pipeline to put it on my phone [16:55] when I'm working on it with my computer. [16:57] The task I gave it here was not trivial. [16:59] I asked it to build an app that uses the [17:01] AR features and the camera on my phone [17:03] to put a hat on my head. And it actually [17:06] works. Like what? It's kind of crazy [17:09] that something like this is able to be [17:12] done just playing around with a web app [17:14] like this. You might notice the hat's [17:15] like in front of my head. That's not [17:18] quite right. I asked it to fix that. [17:20] Let's see if it succeeded. You see [17:22] there's a little building button here in [17:24] the corner and I can tell it to rebuild [17:26] and update the app on my phone directly. [17:28] I also just want to say the onboarding [17:30] is great. All of these types of little [17:31] things that are really annoying to get [17:33] right, it just tells you how to deal [17:35] with them. And I've never had such a [17:36] smooth setup doing an iOS thing in my [17:38] life. Even official Apple stuff is [17:40] significantly harder than this was. Holy [17:42] [ __ ] it actually worked. What? Like [17:46] these are the types of things that just [17:47] were not realistic for people to do if [17:49] you didn't know iOS. Well, you can build [17:51] anything with this. As far as I can [17:53] tell, they've shown off everything from [17:54] subway surfers clones to flight [17:56] management apps to full-on Pokémon Go [17:59] clones. You can use all of the things [18:01] your iPhone is capable of doing, all of [18:03] the APIs, all of the stuff that would [18:04] normally be a blocker. But the biggest [18:06] deal by far, two clicks to App Store. I [18:09] am blown away with Rorm Max, and I bet [18:11] you will be too. Check it out now at [18:12] soyv.link/roor. [18:14] While sponsors can make me money, ads in [18:16] this generic fashion cannot make enough [18:18] money to cover inference. Hell, even my [18:21] sponsors would not be willing to cover a [18:23] dollar plus per prompt for all of you [18:24] guys to prompt for free. Just being [18:26] realistic here. So what other reasons [18:28] would somebody want to do free or [18:30] subsidized inference? Another one that I [18:32] see people talk about a lot and is real [18:34] some amount is the data. We have now [18:36] seen in various different scenarios that [18:39] having chat histories with LMS that have [18:41] verified good results at the end is [18:44] really powerful for training new models. [18:46] You will never be able to make a model [18:48] smarter than the one that generated the [18:49] original data you're training on, but [18:50] you can get goddamn close and you can do [18:52] it a hell of a lot cheaper. This is why [18:54] Enthropic is so scared of all these [18:55] Chinese labs getting access to prompts [18:57] and other data from enthropic models [18:59] that they can use to do their own [19:00] training. This is also why there have [19:02] been some sketchy things that I've heard [19:04] about and wish I had better examples to [19:06] reference and show you guys, but I I [19:07] know from a good friend that there were [19:09] some Chinese companies offering cheaper [19:11] subscriptions for Claude code that would [19:13] require you to use their weird cla [19:15] wrapper thing and it would still use [19:17] anthropic models for a much cheaper [19:19] price like 10 bucks a month I think. And [19:21] the reason they did that is they wanted [19:23] to middleman every request and collect [19:25] all of the inputs and outputs so they [19:27] could use that to go train things. At [19:28] least that's my assumption because I [19:30] don't have any of this on the record. It [19:31] was very illegal. They're also probably [19:33] using all sorts of questionable methods [19:34] to get that inference and all of that. [19:36] But like that type of bootlegging is [19:38] valuable because the data is so [19:39] valuable. I would even argue that in the [19:41] case of something like cursor, they [19:43] benefit greatly from this. obviously [19:45] don't benefit enough to give you free [19:47] inference because to be frank the people [19:49] who are unwilling to pay are also [19:50] working on jack [ __ ] [ __ ] in terms of [19:52] useful stuff. So if you can't afford a [19:55] $20 a month sub your inference is worth [19:57] less than the average person. If you [20:00] don't have a job at a real company [20:01] working in a real codebase and you can't [20:03] afford a cursor sub because of that we [20:05] don't want your data. It's not useful. [20:07] But if you are working at a company and [20:08] you're just trying to minmax what you [20:10] get per dollar, cursor can justify [20:12] subsidizing some amount because if they [20:14] can get the data from your work and from [20:16] your prompting, they can use that to [20:18] train their own better, cheaper models [20:20] that they can then use to lower their [20:22] own costs. That all said, the data is [20:24] not valuable enough to make this worth [20:25] giving out for free. It just helps [20:27] cursor eat a little more cost than they [20:29] would have otherwise been willing or [20:31] able to do. And here's where we get to [20:32] the real reason the companies are doing [20:34] this. Stealing customers. This is a [20:37] desperate attempt from the companies [20:39] that have the money to afford it to [20:41] convince people to use their thing. And [20:43] generally speaking, there's two reasons [20:45] somebody will change the thing they are [20:47] using. It's because there's something [20:48] else that's better or there's something [20:50] else that's good enough and cheaper. [20:51] Those are the two reasons people will [20:53] try something new. In a world where [20:55] everything around us is changing, even [20:57] identifying what is better has gotten [20:59] really hard. And at the same time, a lot [21:01] of these things have gotten more [21:02] expensive. Like I went from having like [21:04] two $20 a month subs at the start of [21:06] last year to three $200 plus dollar a [21:08] month subs. That happened real quick. [21:11] That means getting customers is harder [21:13] cuz I'm not going to just go blindly [21:14] sign up for your $200 a month plan. I [21:16] need to know it's better enough for my [21:18] existing $200 a month plan to even [21:20] consider it or look at it or care. But [21:22] if you offer it to me for free, then I [21:24] might take a look. And if your thing is [21:26] good enough and free, I might go cancel [21:28] the things I'm paying for and stick with [21:30] yours. or if it's cheap enough, just [21:32] move to it and be happy. If you can't be [21:34] better than your competition or you're [21:36] struggling to communicate the ways in [21:37] which you're better than your [21:38] competition, price becomes the easiest [21:40] way to communicate. And I'm definitely [21:43] not speaking from experience here. [21:44] Putting out the $8 a month chat app, T3 [21:47] Chat, definitely not something I would [21:48] know way too much about. If you can't [21:50] convince people you're better, convince [21:52] them you're cheaper cuz that's easy. And [21:53] then when they move over, they will [21:55] discover how much better what you built [21:57] is if it's better. Or they won't. So [22:00] while going cheap or free can give you a [22:02] huge boost in potential new users, if [22:05] you're not good enough, you will not [22:07] keep them and you will just be wasting [22:09] money on the people who are leftover. So [22:12] if we have this range of people and on [22:14] the right side we have billionaires and [22:17] on the left side we have 12 year olds [22:20] learning to code, the spend potential [22:22] for these groups is massively different. [22:25] And if you can make a product that is [22:27] beloved by billionaires, you should [22:28] probably do that because you can make a [22:30] lot of money. But there's a line [22:31] somewhere in this spectrum that is I'll [22:34] try it if it is free. And this line is [22:37] the point at which people will be more [22:39] interested in a thing if it's given out [22:41] for free. This line's probably further [22:42] to the right here. Even like some of the [22:44] rich people I know are much more [22:45] interested in trying a free thing than a [22:47] paid thing. So there's a line somewhere [22:50] here. But there's also an important [22:52] range to identify here, which is the [22:54] range of I will pay for the better [22:56] thing. So if you make your thing free [22:58] and you want to get paying users long [23:01] term, this range here is a gold mine. [23:04] The range between I'll try it if it's [23:06] free and I'll pay for it if it's better. [23:08] If you can get them in cuz it's free and [23:10] then they stick with it cuz it's better [23:11] and they start spending money, awesome. [23:13] You can get a lot of good users from [23:15] this bucket. I would argue something [23:16] like GitHub has done this really well [23:18] where GitHub is a thing we all use [23:20] because it's free and we see everyone [23:21] using it and then when we go to a job we [23:23] want to use it there. We start paying [23:24] and they make a lot of money. But here's [23:26] the problem. We are expecting that users [23:28] in this area will be happy enough with [23:31] the product that they become paying [23:33] users. But what if the product isn't [23:35] very good? Well, then we hit a different [23:37] line that exists in this section. I can [23:40] only use it if it's free. And this line [23:42] is a death sentence. Everything left of [23:45] here is a suicide mission. If your only [23:48] users are people who are using your [23:50] thing because they can't afford to use [23:51] the better thing, you're [ __ ] You're [23:54] absolutely [ __ ] You will never [23:56] recover from this position. The only [23:57] reason so many of these companies [23:59] offered free tiers in the past, like [24:00] preai, is because everyone would try it [24:03] if it's free and enough people would [24:04] start paying that it was worth it. But [24:06] if all the people over this line are [24:08] happy and aren't going to try your [24:09] thing, all you're getting by making it [24:11] free is a bunch of [ __ ] users that are [24:14] going to cost you a ton of money and [24:15] never convert into real paying users. By [24:18] being [ __ ] and being free, Google has [24:21] effectively filtered down the [24:23] anti-gravity user base to be early [24:26] programmers who don't really know what [24:29] they're doing and won't be valuable and [24:31] won't spend money or really experienced [24:33] programmers who won't spend money [24:35] either. looking at you Linus Tovals. I [24:37] saw you vibe coding in anti-gravity for [24:39] some reason because it was free. But [24:41] that's the reality of this. Most of [24:43] these companies are doing the free thing [24:45] because they're hoping someone like [24:46] Lionus Torvos will try it because it's [24:48] free. But in reality, what they end up [24:50] with is all of these people who are only [24:52] using it because they can't afford [24:54] anything else. And those users are going [24:55] to waste a bunch of your GPUs and energy [24:57] and time and support. Especially these [25:00] cheap and free users are so expensive [25:02] support-wise because they have so many [25:04] needs. There's the classic meme of the [25:06] $100 client will ask for all of your [25:08] time and attention. The $1,000 client [25:11] says looks good. That's a very real [25:13] thing. The users in this range aren't as [25:16] expensive as the paying users or less [25:18] expensive. They are often more [25:20] expensive. They have higher [25:21] expectations. They are more annoying. [25:23] And the likelihood they ever make you [25:24] money isn't near zero. It is zero. And [25:27] Google's inadvertently put themselves in [25:29] a position where the majority of the [25:31] users of their code stuff live in this [25:33] cursed corner. And that's why Google has [25:35] to cut things off. But I want to compare [25:37] this to other businesses. I want to [25:39] compare this to something like Claude [25:40] Code. There's another line somewhere in [25:42] this that is I can afford $200 a month. [25:45] That doesn't mean you're willing to just [25:46] go sign up and use it for that, but you [25:49] are capable of spending that much money. [25:50] So maybe you tried Claude on the free [25:52] tier on the website. Maybe you got one [25:54] of those guest pass things and you're [25:56] trying out cloud code. You're like, [25:57] "Yeah, this is pretty impressive. I can [25:58] see myself paying for it." So, you bump [26:00] to the $20 tier. You use all your usage. [26:02] You're like, "Yeah, this is useful. I'll [26:03] pay a bit more." You bump up again and [26:05] then eventually you're suddenly at the [26:06] $200 a month tier. You then see that [26:08] Open AAI put out a new model and you [26:10] really want to use it. You don't want to [26:12] have a second $200 a month sub though. [26:14] So, instead, you go try out Cursor [26:17] because Cursor has a $200 a month plan, [26:19] but it has Claude and it has OpenAI [26:21] models in it. you were barely even using [26:23] half your quad sub anyways, so you'll be [26:25] fine. So you move over to cursor and you [26:27] hit your limits in the first 3 days and [26:30] you're like, "What the [ __ ] I didn't [26:31] even get close to hitting my limits when [26:33] I was using the Claude plan. Why is the [26:35] same money I'm spending on Cursor [26:37] hitting limits so fast? Did I use the [26:40] wrong model? Did I do something bad? Is [26:41] Cursor just overcharging cuz they're [26:44] greedy?" No, it's actually quite the [26:45] opposite. Turns out that that $200 a [26:47] month plan can be used for up to $5,000 [26:51] of compute per month. So why would they [26:53] ever do this? Let's think about some [26:56] user archetypes. Let's focus on that [26:58] 12-year-old again. 12-year-old could use [27:01] up to 15 in inference. They pay $0. They [27:05] will pay zero over lifetime. This user [27:08] is not going to make the company money. [27:10] And there are millions of them. This is [27:12] not a very appealing customer breakdown. [27:15] This is not the person I would want to [27:17] have as my user. So, what about some [27:20] 42year-old dev? They could use up to [27:23] $5,000 in inference. They pay 200 a [27:26] month. They might pay 1 mil plus over [27:29] lifetime or get their company to. And [27:32] there's a few hundred,000 at most of [27:35] these people. So, on one hand, this [27:37] 42-year-old dev is costing Anthropic way [27:40] more than this 12-year-old is. The [27:42] 12-year-old's costing 15 over zero. The [27:44] 42y old dev is costing five grand over [27:46] 200. So they're costing Anthropic [27:48] theoretically up to $4,800. And the [27:51] 12-year-old's only costing them 15. But [27:52] let's think about this longer term. What [27:54] if the cost of inference goes down by [27:57] 10x? This free user goes down to $1.50, [28:01] but that's not enough to be covered on [28:02] ads. That's still not profitable. And [28:04] they'll never spend money. So this user [28:06] still sucks. But that $5,000 of [28:08] inference is suddenly $500. Obviously [28:12] still not profitable. they're still [28:13] losing 300 bucks on that. But if the [28:15] price gets 10x cheaper, it gets a hell [28:17] of a lot closer. It also is worth noting [28:19] that most users don't come close to [28:21] using all of that inference. So if your [28:23] top 1% user is using the full amount [28:26] they're allowed, but the mean users [28:28] using less than half, you're now [28:30] profitable with this change easily. But [28:32] if it gets 100 times cheaper, the [28:34] 12-year-old still cost 15, which might [28:36] be almost enough to cover them with ads, [28:39] or at least close. But it also means [28:41] that Anthropics cost is 50 bucks and now [28:43] they're getting a 4x profit on a top [28:46] tier user. This is just one of the [28:48] different dimensions that has to be [28:49] considered here. The more important one [28:51] is that if you give me a $200 a month [28:52] plan, I go use it personally. I really [28:54] like it. And then I go to my company and [28:56] say, "Hey, we should use this here." [28:57] They subscribe the entire engineering [29:00] org to it. 20% of that org actually uses [29:03] it. And 10% of them use it heavily. [29:05] That's 80% of my spend going to people [29:08] barely using the thing. Just paying free [29:10] money to Anthropic. I'm kind of doing [29:12] this myself because I'm maintaining the [29:13] $200 a month sub and I barely use the [29:15] cloud code one because I'm too busy [29:16] using things like cursor and codec cli [29:19] mostly codec cli 33 code. So I'm barely [29:21] using my inference and that helps [29:23] subsidize the cost for others. But if [29:24] all of your users spend $0, they can't [29:26] subsidize each other. And to those who [29:28] are saying, "But what about the kids [29:30] when they grow up? They might use the [29:32] thing when they're adults. It's not that [29:34] they're going to be 12-year-old and poor [29:35] forever. Yeah, they're going to become [29:37] adults and they're going to move to a [29:38] good thing. If you were right about this [29:40] assumption, then we'd have a bunch of [29:41] 22-year-old Genzers who are using [29:44] Chromebooks as their daily driver [29:45] computer. You're just wrong. When the [29:48] kids grow up, they get better things. [29:49] And the only reason it makes economic [29:51] sense to make things for the poor kids [29:53] is if you can make money off of them. [29:55] And you can't. And you're not going to [29:56] make that money later. You're just [29:58] wrong. And thankfully, only one company [30:00] was stupid enough to think otherwise. [30:02] And it's everyone's favorite Google. And [30:04] Google has now realized how badly they [30:06] [ __ ] themselves by doing this. And [30:07] they're trying to lock it down. I will [30:09] say with near 100% certainty that this [30:12] type of subsidization has to die. There [30:15] is no reason this ever should have [30:16] existed. It was bad math on some PM's [30:18] [ __ ] plate that resulted in them [30:21] giving out way too much inference for [30:23] free or really cheap in a desperate [30:25] attempt to get somebody to care about [30:27] Google's developer products. And they [30:29] failed outright. This type of [30:31] subsidization is a little bit more [30:33] interesting. On one hand, it is absurd. [30:35] The amount of infants we're getting here [30:37] is crazy. This can't last. They can't [30:40] possibly afford to do this long term. [30:42] But on the other hand, the amount [30:43] they're subsidizing is going up. [30:45] According to the same reason I cited [30:46] earlier, which was Curser doing their [30:47] own internal testing, last year, you [30:49] could only use up to $2,000 a month for [30:52] the $200 a month plan. So, it was a 10x [30:54] subsidization. It is now a 25x since [30:57] then. Anthropic is increasing how much [31:00] you're getting in terms of a discount [31:01] when you use their models. And there's a [31:03] lot of reasons and a lot of incentives [31:04] they have to do that. But the biggest is [31:05] that they want to get people off of [31:07] cursor where every model can be used and [31:10] onto claude code where only enthropic [31:12] models can be used. They have a very [31:13] strong economic incentive to do this and [31:15] they are going to keep being [ __ ] [31:17] because of it. It's the same reason [31:18] they're being [ __ ] to DAX and open [31:20] code. There's a lot of layers to this [31:21] one, but one of them is that if you use [31:23] other harnesses, it's easier to switch [31:25] to other models. So you're less stuck in [31:28] the anthropic ecosystem. And it does not [31:30] make sense for them to give you $5,000 [31:32] of inference for 200 bucks a month if [31:34] you've moved off before inference gets [31:36] cheaper. These subsidies only make sense [31:38] if they make you a lifetime customer. So [31:40] anything that risks losing you as a [31:42] lifetime customer is a massive risk [31:45] they're not interested in taking. So [31:47] that's why Anthropic is working so hard [31:49] to kick you out of tools that give you [31:51] choice. And this is the only reason I'm [31:53] scared about the cloud implementation in [31:55] T3 code because according to their terms [31:57] of service, their SDKs, their own [31:59] guidelines, the advice I've gotten from [32:01] existing employees and people online, [32:03] and most importantly, according to [32:04] Claude, our implementation is totally [32:07] fine and safe. But the economic reality [32:09] of Anthropic not wanting an option that [32:12] gives you choices because then their [32:14] subsidies don't guarantee profit long [32:17] term, that's a real risk. So the only [32:20] reason T3 Code would end up not having [32:22] good cloud code support over time is [32:24] because Anthropic specifically doesn't [32:27] want you to have options. So why does [32:29] OpenAI do subsidies then? I think [32:31] they're getting like the opposite effect [32:33] because some people including this [32:35] particular dev have switched off of [32:36] Claude because they're tired of the [32:38] model just being shitty in favor of [32:40] something like OpenAI's models. This is [32:42] largely because of their growth phase. [32:44] In fact, I can show you. Try the Codex [32:47] app with 2x rate limits until April 2nd. [32:50] By the way, it is my understanding that [32:51] these limits apply in all Codex [32:53] services, not just the Codeex app. But [32:56] the 2x rate limits they're giving you [32:58] here are because they are trying really [33:00] hard to pull people over from other [33:02] options. Because Codeex is significantly [33:05] less popular than Cloud Code still to [33:07] this day, they had to do whatever they [33:09] could to try and pull users over. And [33:11] these rate limits are too generous. I [33:13] haven't even hit more than 30% usage [33:15] when I was trying. One of my employees [33:18] got close to 90% once and it was Julius [33:21] and he was trying [ __ ] hard to do it. [33:23] And then there's the reset button that [33:24] Tibo keeps hitting. That's a whole [33:26] separate set of jokes. The reason [33:27] opening I can do this and justify it is [33:29] because they have so many subscribers on [33:32] the $20 a month tier doing nothing that [33:35] it more than justifies this. it as a [33:37] land grab, as an opportunity to try and [33:39] take some ground from Claude code is [33:43] worth the money here. But again, to go [33:45] back to my list, this is all part of [33:47] three. The only strategy that makes real [33:49] sense. The reason to give away free and [33:51] subsidized inference is to steal [33:53] customers. It will not create new [33:55] customers. It will not grow the pie. The [33:58] reason you do this is to take more of [33:59] the pie away from someone else. And [34:01] that's why OpenAI is doing a lot of [34:03] different things now. Not only is this [34:04] why they're giving us the temporary 2x [34:06] rate limit, this is also why they're [34:08] working with people like open code and [34:09] like us with T3 code to make sure we [34:11] have everything we need to integrate [34:13] because they don't want to be the only [34:14] option. They want to be the best option [34:17] because right now they are the option [34:18] with 30% of the market at most. But if [34:21] they give everybody a list of options [34:23] and theirs is the best value and the [34:25] best option, they will start to win more [34:27] and more. And this is going very well [34:29] for them so far. Whereas Enthropic while [34:31] having incentive to subsidize to keep [34:33] their holds, they do not have incentives [34:36] to keep offering their product near [34:38] others. This is why you'll hear [34:39] engineers working on Codeex talk about [34:41] Claude meaningfully, but you'll really [34:44] struggle to get a Claude engineer to [34:45] even mention Codeex. It is just the [34:48] unique shape of reality distortion [34:50] bubbles they both live in. But it all [34:52] comes down to where they are in the [34:53] market. And if I'm being honest, as much [34:55] as I've loved working with OpenAI and as [34:57] great as the people I'm friends with [34:58] there have been and just are pleasant to [35:00] work with, I would suspect that if they [35:02] get over 50% market share, that their [35:04] tune will change quickly. Much like it [35:06] has for Anthropic, who went from [35:07] reasonable, somewhat generous to if [35:10] you're not using cloud code, go [ __ ] [35:12] yourself. And that happened really fast. [35:14] So where does Google fit in all here? [35:17] They don't because they're clueless. [35:19] Google as a business is desperately [35:20] trying to get real AI customers. And [35:22] they're doing that by forming dozens of [35:24] different teams across the company with [35:26] no communication and understanding of [35:27] each other that are fighting over GPUs [35:29] and allocation. And I've known people [35:31] who work on dev tools at Google that [35:34] could not convince Google to let them [35:36] add Gemini Pro 3.1 because they wanted [35:38] to reserve it to give it to the free [35:40] users in anti-gravity. [35:42] Yeah, Google subsidized so hard they [35:44] [ __ ] over their own developers. And [35:46] it's truly insane how much things are on [35:49] fire there as a result. where Enthropic [35:52] does these types of things with a very [35:53] focused strategy, Google's doing them [35:55] because they're incompetent. They [35:57] accidentally put themselves in a bad [35:58] position. So, while I'm doing my best to [36:00] give you guys this framing that can help [36:01] you understand all of these companies [36:03] and why they work the way they work, [36:05] none of these framings will ever work [36:06] for Google because Google is not working [36:08] based on what makes sense. Google is [36:10] working based on how Google works, which [36:12] is not particularly well. I've been [36:14] planning on doing the Gemini is bad [36:16] vibes video for a while now. I don't [36:17] even know if I'm going to bother because [36:19] I don't think anyone cares. The only [36:21] people I knew who were using the Gemini [36:23] models and were using things like [36:24] anti-gravity heavily were the people who [36:26] couldn't afford other things. And now [36:28] that they have officially been booted [36:29] from doing such, I suspect no one will [36:32] be asking me going forward. Hopefully, [36:33] this helps you in your understanding of [36:35] how these business incentives work, and [36:37] maybe you'll look at your $200 a month [36:39] sub with a little bit less scrutiny. or [36:40] at the very least you'll be a little bit [36:42] more generous to services like us with [36:44] T3 chat that have to pay the full price [36:46] when nobody else does. It's a great time [36:48] to be a user of these subscriptions, but [36:50] it is a terrible time to be a competitor [36:51] trying to fight them. That's why all [36:52] these companies are doing it. That's why [36:54] small startups like ourselves get kind [36:55] of screwed. And that's why you, a [36:57] developer building with these tools, [36:58] have an incredible opportunity to do [37:00] awesome things. Seriously, if you can [37:02] afford one of these subs and you don't [37:03] have it yet, you really should go find [37:04] one. even if that means not using [37:06] something like what we built with T3 [37:08] chat because they are such a high value [37:10] return if you can actually use them that [37:12] it's one of your last chances to get [37:14] this type of free value in life. You [37:17] should take advantage of it while you [37:18] can, but expect it to disappear any day [37:20] as we're starting to see now. And [37:21] hopefully prices will get cheap enough [37:23] that we don't need this long term. Until [37:25] then though, I'm going to do my best to [37:26] max out my $200 a month subs. You should [37:28] too. And if you're looking for a place [37:29] to do it, you should try out T3 Code. We [37:31] offer all of these things in an open [37:32] source rapper. Anyways, I have nothing [37:34] else on this one. I'm curious how y'all [37:36] feel.  📄 **Full text:** ![[Full Document Contents/Articles/AI Has a Subsidization Problem|AI Has a Subsidization Problem, YouTube]] ## Highlights - their own models and making their own whatever they're doing and the end users who are paying the whatever a month to go use the models in products like anti-gravity or T3 code or just using it in their chat interface directly. And since Google in particular desperately wants to steal customers off of OpenAI and off of Anthropic, they've been way too generous with subsidizing usage in their products. I cannot tell you how many times I have seen comments from lower income people in my content, be it ([View Highlight](https://read.readwise.io/read/01kmjqq9b6fhx1b1tsaazkjrrm)) - Tags: 👻 ai highlighted - Twitter, Twitch, YouTube, or wherever, saying that I need to stop being so mean to anti-gravity because it's the best option for poor people who can't afford a subscription for something better. to which I would usually bite my tongue because I know the reality here. Those customers aren't valuable enough to get free inference. There is no world in which that continues. And we are now at the end of that. I had a feeling it would come, but I honestly thought Google's billions upon billions of dollars would sustain this a bit longer. Google got users through giving out free ([View Highlight](https://read.readwise.io/read/01kmjqq9bybsh036jycny41j9j)) - Tags: 👻 ai highlighted - inference, but that meant the majority of their users were using free inference, which meant that they were eating a ton of costs and had a lot of GPUs that were being used by users who would never spend actual money when they could be used by people who are currently spending actual money. And this has affected us. We've had times where people are using paid things on T3 Chat where we're paying API prices and they send a message and 3.1 just fails to respond because it's overloaded. or even I on the $250 a month Google plan that I subscribed to forever ago for ([View Highlight](https://read.readwise.io/read/01kmjqq9d9d4z0w871wmj2wd1w)) - Tags: 👻 ai highlighted - content and probably should unsubscribe, but it's fun to have it there to [ __ ] about things like this. I couldn't use the new Gemini 3.1 Pro model when it dropped in the $250 a month subscription and the Gemini CLI. And when asked, they cited this as a capacity issue because they're giving it away for free too much. There is no such thing as free inference. Either you're the product and you won't be for long or something illegal is happening. So, if you were picking these tools because they're free and you can't afford it and you didn't use the two months of thousands of ([View Highlight](https://read.readwise.io/read/01kmjqq9eax3etyf4r4xpmwc8z)) - Tags: 👻 ai highlighted - preai, is because everyone would try it if it's free and enough people would start paying that it was worth it. But if all the people over this line are happy and aren't going to try your thing, all you're getting by making it free is a bunch of [ __ ] users that are going to cost you a ton of money and never convert into real paying users. By being [ __ ] and being free, Google has effectively filtered down the anti-gravity user base to be early programmers who don't really know what they're doing and won't be valuable and ([View Highlight](https://read.readwise.io/read/01kmjqq9he927rgm69qnq8gbyv)) - Tags: 👻 ai highlighted - users of their code stuff live in this cursed corner. And that's why Google has to cut things off. But I want to compare this to other businesses. I want to compare this to something like Claude Code. There's another line somewhere in this that is I can afford $200 a month. That doesn't mean you're willing to just go sign up and use it for that, but you are capable of spending that much money. So maybe you tried Claude on the free tier on the website. Maybe you got one of those guest pass things and you're trying out cloud code. You're like, "Yeah, this is pretty impressive. I can see myself paying for it." So, you bump to the $20 tier. You use all your usage. ([View Highlight](https://read.readwise.io/read/01kmjqq9p165y36n4m8hxj6qtj)) - Tags: 👻 ai highlighted - org to it. 20% of that org actually uses it. And 10% of them use it heavily. That's 80% of my spend going to people barely using the thing. Just paying free money to Anthropic. I'm kind of doing this myself because I'm maintaining the $200 a month sub and I barely use the cloud code one because I'm too busy using things like cursor and codec cli mostly codec cli 33 code. So I'm barely using my inference and that helps subsidize the cost for others. But if all of your users spend $0, they can't subsidize each other. And to those who are saying, "But what about the kids ([View Highlight](https://read.readwise.io/read/01kmjqq9yqax9e2748cd3z76tm)) - Tags: 👻 ai highlighted