Key insights
- This analysis presents a bearish outlook on AI hardware demand, arguing that the commoditization of AI models via open-source alternatives will undermine the pricing power of closed-source providers like OpenAI. This unsustainable business model, reliant on continuous funding and IPOs, suggests that peak capital raises for these AI labs will coincide with the peak of GPU and memory hardware demand. The author predicts a subsequent downturn in the memory upcycle as the market realizes the lack of profitability in the AI software layer, impacting companies like Micron.
Current Memory upcycle could turn into downcycle
I've been staring at the AI supply chain for months and something just doesn't add up. Everyone is consensus long on NVIDIA, TSMC, and the memory suppliers (SK Hynix/Samsung) based on the "insatiable demand for compute" narrative. But the bear case isn't about whether AI is useful—it’s about the fundamental business model failure at the software layer and how that’s about to destroy hardware demand.
Think about it this way:
Enterprises are burning through their API credits 10x faster than they budgeted. I’m hearing that some tech companies torched their entire 2026 AI budgets in less than four months. Businesses cannot afford to keep paying metered API rates at this scale.
The consensus thinks this means OpenAI or Anthropic will just raise prices. They can't.
Why? Because open-source aggregates like OpenRouter, Venice, and Baseten have destroyed their pricing power. It’s the "Food Delivery App" effect. OpenRouter lets a developer switch from a closed-source model to an open-source model with one line of code based on whoever is cheapest that exact hour.
With Chinese labs aggressively dumping frontier-grade open-source models (like DeepSeek V4) at 1/30th or 1/100th the price of closed models, the intelligence layer is being entirely commoditized. The labs get the model for free, and inference providers just charge for the bare electricity and server time.
The closed labs are running on deep negative margins (OpenAI reportedly near -122%). They have zero organic cash flow. This means they are 100% dependent on continuous mega-round fundings and upcoming IPOs just to keep buying GPUs and subsidizing usage.
Here is the kicker for the hardware trade:
The exact moment these labs hit their peak capital raises or IPOs later this year, that will mark the historical peak of hardware demand. The markets will realize these labs are just a bridge to nowhere with no path to profitability. The moment the venture/equity funding stops, the massive Capex cycle reverses instantly.
Even worse for the memory and chip guys: Open-source inference doesn't need massive monolithic clusters. It’s highly distributed and efficiency-driven. Models are being distilled and quantized to run on smaller, cheaper setups. As the market shifts heavily from massive "Training" clusters to cheap "Inference" provided by open-source players, they aren't going to buy top-tier premium HBM configurations. They will hack together custom ASICs or NPU setups with cheaper, high-density standard DDR5 or LPDDR configurations to save on TCO (Total Cost of Ownership).
The call volume (Q) of AI queries might explode, but the dollar value of memory/silicon per server (P $\times$ Q) is going to collapse because the open-source ecosystem forces everyone to build for ultra-low cost.
Change my mind, but the hardware cycle is peaking right now on artificial, venture-subsidized demand. And the culprit in on the demand side, not supply side. When the funding drying up meets the commoditization of inference, the hardware unwinding is going to be brutal. Also read the following tweet where i developed my idea from:
Credit to: https://x.com/Shaughnessy119/status/2061807418829578413