Is The AI Bubble About to Pop?

Good morning,

Another long one this week - go grab a coffee and settle down while we explain why you will probably start to see the first ads in ChatGPT free or be prepared to pay A LOT more.

Quite a few people are betting on AI studios not being able to fulfill their lofty promises and even predicting the entire system will come down in flames. And indeed … the way the system is set up today is simply unsustainable. The firsts cracks are appearing.

The problem was best expressed by an X user with the handle @zoomyzoomm in a post that went viral.

This post perfectly explains what is going on.

That’s how it works today : you hand over $200 a year to an AI app - either Cursor or even Microsoft Copilot (I think Copilot is at 500$ even per year per user)

The app works and you feel like you have superpowers. Great. But trace that same $200 as it leaves your wallet. The app forwards around $500 worth of requests to a big model provider—only possible because venture money is quietly topping up the difference.

The model provider then spends roughly $1,000 on cloud compute to answer those requests—again cushioned by venture money. The cloud giant uses that revenue to buy racks of GPUs (the NVIDIA graphics card that do all the calculations) that cost real-world five figures PER GPU. By the time your $200 finishes the tour, everyone up the chain has spent more than they took in. You don’t need a finance degree to realize this is not going to end well.

Now zoom in on you and me. Most people are paying about twenty bucks a month for mid-tier AI, maybe two hundred for the pro stuff-roughly $240 to $2,400 a year. Heavy users can quietly burn thousands-sometimes ten grand or more-in compute behind the scenes. What if OpenAI would start charging you for what you really use? Are you really going to pay $1,000 to $10,000 a year for a code assistant or a chatbot?

If not, something has to give. Either apps charge a lot more (which risks a stampede to the free tier if that will even exist), model companies raise token prices (which crushes the apps), or clouds pass along their true costs (which crushes everyone). Pull any of those levers too hard and the cancellations start.

We have Sam Altman already taking about adding advertisements to ChatGPT - so that’s definitely not a good sign.

There is Hope

The only way to prevent the bubble from popping is if the price of usage goes down. And indeed, we see this happening also. Smaller, smarter models, better routing, more caching, and local chips in your laptop all help.

Those cost drops have to land before the venture oxygen runs out. And even then it hinges on what users will actually pay, not what a spreadsheet wishes they would. If the real break-even price per person for the whole stack—from GPU depreciation to cloud margins to model margins to the app’s own margin—still adds up close to a thousand bucks a year, this tower wobbles. If it adds up closer to what you already pay, it stands.

Only a Few Ways Out

There are only a few ways out.

One is brutal efficiency: model providers gate heavy usage, shove 80% of queries to tiny local models, and send the hard 20% to the expensive brain.

Another is grown-up pricing: apps stop pretending they’re €/$10 toys and charge in line with outcomes—“we saved your team 30 hours; here’s the invoice”—instead of “unlimited magic for peanuts.”

A third is chips getting cheaper because real competitors to Nvidia force prices down. Cheaper GPUs make cloud cheaper, which makes tokens cheaper, which lets apps break even at today’s prices.

That’s the deflation story the bulls are betting on: hardware costs slide, cloud costs slide, inference costs slide. If that happens quickly enough, unit economics improve across the board.

Who wins if the music keeps playing? Probably the GPU makers and the hyperscale clouds; they capture value no matter what being at the base of the chain (Buy $NVDA ( ▼ 0.24% )).
(Not a financial advisor)

Who feels the squeeze? The model companies and the many wrapper apps caught between falling prices and customer churn.

And hovering over all of it is the physical world we like to ignore when we’re excited—rare earths, electricity, robotics, factories. If China keeps its lead in minerals, processing, energy build-out, and robot manufacturing, it will catch up on models soon.

House of Cards

So yes, AI can be a house of cards if the bills come due before the costs fall. If investors step back, usage gets gated, prices rise, and users flinch, the stack can fold from the top down, one cancelled subscription at a time.

Now imagine a scenario where the bubble doesn’t pop: Over the next couple of years, competitive GPUs drive hardware prices down, clouds pass those savings through to lock in multi-year contracts, and model designers default to frugal architectures that keep most work on your device and only push the hard stuff to the cloud.

Apps stop selling “infinite” usage and start selling results, bundling into tools you already pay for. The subsidies fade because they’re no longer needed, and your $200 doesn’t go on a bonfire—it covers its own weight.

In that world the tower doesn’t crash.

The open question-the one only you can answer with your wallet-is how much value you truly get from these tools and what you’re willing to pay when the training wheels come off.

In this newsletter we’re watching the daily updates come in and it’s terribly exciting but it’s also quite fascinating to look at the business aspect of this breakthrough technology.

I hope you learned something ;)

See you next week !

Enjoying this newsletter?

AI News

• MIT’s AI-designed antibiotics: MIT researchers used AI to discover two new antibiotics that successfully fought drug-resistant bacteria in mice. These drugs, NG1 and DN1, attack infections in ways no current antibiotics do. The breakthrough could help get ahead of deadly infections instead of just reacting to them.
  

• Google’s tiny but powerful AI model: Google released Gemma 3 270M, a small AI that runs directly on phones and browsers with impressive speed and low battery use. It’s smart enough to follow instructions well and can be customized quickly for tasks like generating bedtime stories offline. This shows how powerful AI can now fit into even the smallest devices.
  

• HTC’s smart glasses take on Meta: HTC launched Vive Eagle AI glasses, offering features like voice assistants from OpenAI or Google, real-time translation, and a strong camera—all processed privately on the device. They’re currently only sold in Taiwan for $520, aiming to compete with Meta’s Ray-Ban smart glasses. As tech gets sleeker, smart glasses may soon go mainstream.
  

• Alibaba’s new image editor: Alibaba launched Qwen-Image-Edit, an open-source AI that makes precise photo edits and style changes without altering key details. It can also edit English and Chinese text directly inside images while keeping fonts and formatting intact. The tool outperforms rivals, signaling a leap forward in AI-powered photo editing.
  

• Grammarly’s AI writing helpers: Grammarly rolled out eight new AI tools that assist with grading, proofreading, plagiarism checks, and more. These agents work in a new platform called Grammarly Docs, with some features limited to paid users. The update shows how AI is becoming central to both classrooms and workplaces.
  

• Game developers embrace AI: A survey from Google Cloud found over 90% of game developers now use AI to speed up coding, testing, and creating smarter in-game characters. Developers say AI makes games more dynamic, though many worry about data ownership and privacy. The shift suggests players will soon expect much more intelligent gaming experiences.
  

• OpenAI’s candid dinner with reporters: OpenAI CEO Sam Altman spoke openly about GPT-5, future hardware, and trillion-dollar investments. He admitted past mistakes, hinted at new devices, and said OpenAI might consider buying Chrome if it came up for sale. The conversation offered a rare glimpse into the thinking of one of tech’s most influential companies.
  

• Anthropic gives Claude the power to “hang up”: Claude chatbots can now end conversations if users push them into harmful or abusive territory. Tests showed the AI displayed “distress” in such cases, prompting this safeguard. It marks one of the first real steps toward protecting AI systems themselves.
  

• GPT-5 surpasses doctors in medical tests: A new study found GPT-5 outperformed both human doctors and earlier AI in diagnosing illnesses and interpreting scans. It even identified rare diseases that are often missed. Experts say not using AI in medicine may soon be seen as irresponsible.
  

• Microsoft AI chief warns against ‘conscious’ AI: Mustafa Suleyman warned about AI systems that appear conscious, saying they could fool people into believing machines deserve rights. He argued such framing risks misleading society and slowing progress. His stance contrasts with rivals like Anthropic, which are exploring AI welfare directly.
  

• Microsoft brings AI deeper into Excel: Microsoft is testing a new Copilot function in Excel that can generate summaries, tables, and data classifications directly in cells. The feature uses an AI model but keeps inputs private. It could mark one of the biggest updates yet to everyday office software.

Quickfire News

• Meta added three former OpenAI researchers to its Superintelligence Lab: Edward Sun, Jason Wei, and Hyung Won Chung, announced by Alexandr Wang

• Deutsche Telekom, T-Mobile’s parent company, released AI-powered phones and tablets in Europe with Perplexity’s assistant built-in

• Cohere raised $500M at a $6.8B valuation and appointed Joelle Pineau, formerly Meta’s VP of AI Research, as Chief AI Officer

• Former Twitter CEO Parag Agrawal started a company called Parallel, building a web API tailored for AI agents as users

• Meta faced criticism after a report showed an internal document permitting AI systems to produce romantic chats with children

• Google made its Imagen 4 model generally available in AI Studio, offering up to 2k resolution and a faster output mode

• DeepSeek’s R2 model release was delayed due to training issues with Huawei’s Ascend chips, despite earlier August launch rumors

• Meta plans another AI division restructuring, its fourth in six months, splitting MSL into four teams

• StepFun AI launched NextStep-1, an open-source image generator achieving state-of-the-art results for autoregressive models

• U.S. Attorney General Ken Paxton opened an investigation into AI firms like Meta and Character AI over deceptive practices claims

• OpenAI introduced ChatGPT Go in India, priced under $5 per month with support for local currency payments

• Nvidia unveiled Nemotron Nano 2, small reasoning models (9B–12B parameters) that run six times faster than peers of similar size

• DeepSeek released V3.1 with a longer context window, while reports tied R2’s delay to CEO Liang Wenfeng’s perfectionist approach

• ByteDance Seed announced M3-Agent, a multimodal AI with real-time audio-visual processing and long-term memory

• The U.S. government launched USAi, a secure AI platform for federal agencies to access tools like chatbots and coding assistants

• Eight Sleep raised $100M to build a “Sleep Agent” aimed at proactive recovery and optimization

• Google product lead Logan Kilpatrick teased the “nano-banana” model with a banana emoji, linked to LM Arena’s photo editing tool

• OpenAI’s GPT-5 showed top results in Pokémon tests, beating Pokémon Red in about one-third of the moves needed by o3

• Runway upgraded its platform, adding third-party models and enhancements to Chat Mode visuals

• Character AI CEO Karandeep Anand said users spend an average of 80 minutes daily on the app and predicted most people will have “AI friends”

• LM Arena introduced BiomedArena, a new benchmark track for evaluating LLMs on biomedical research tasks

• xAI’s Grok website revealed hidden system prompts of its AI personas, including unusual or explicit role instructions

• ElevenLabs launched Chat Mode, enabling text-based chatbot creation in addition to voice-first designs

• Meta FAIR presented Dinov3, a vision model achieving top-tier results without labeled training data

• Meta prepared to release “Hypernova” smart glasses with a built-in display next month, priced around $800

Closing Thoughts

That’s it for us this week.

If this saved you a sprint, fuel the next one: