Let's go back to the starting point of this major crash and see what actually happened. Was it a bubble burst or excessive panic? A correction or the end?

On July 1, a headline news story said Meta might sell its excess AI computing power, creating a business model similar to NeoCloud. The market interpreted this as the first real signal of "excess computing supply." Combined with earlier concerns about memory accounting for too high a share of total CapEx, a systemic sell-off began.

1️⃣META LOST IN THE AI RACE
This should already be an open secret. Beyond lagging behind in models, massive AI investments have further drained Meta's cash flow. Without any changes, Meta's free cash flow will remain negative for years to come.

Currently, the utilization rate of H100/H200 computing power within Meta's internal infrastructure is about 65%, leaving 35% idle capacity as a monetization path for cash-strapped Meta.

That's no small amount of money.

2️⃣TIERS OF COMPUTING POWER
Meta mainly leases out GPU clusters from the H100/H200 generation. The latest "top-tier training compute GB300" is still primarily for its own use.

Meta currently plans two models for leasing computing power:
1) Lease raw compute, allowing customers to train/infer on Meta DC (similar to CoreWeave);
2) Open access to AI models hosted on Meta's infrastructure.

Demand for inference cards vs. training cards will diverge, with old cards used for inference and new cards leading training as the main trend.

Top-tier training compute remains scarce. Delivery lead times for high-end training compute are still 6–9 months or more.

3️⃣IS AI DEMAND SLOWING DOWN?
SemiAnalysis provided specific numbers: In the first half of 2026 alone, Meta has already signed contracts for over 5GW of data center capacity, including cloud leases and colocation facilities, and this does not include all the progress of self-built projects.

Earlier it was mentioned that 35% of computing power is idle, so why keep buying new computing power?

The meta Superintelligence Lab (MSL) lists large model R&D as the top priority for computing power usage, supporting the training iterations of the next Llama series and multimodal models in an attempt to catch up with OpenAI/Anthropic.

Ad recommendation system (RecSys): 10x expansion space.

SemiAnalysis believes that Meta thinks it can expand the complexity of its ad recommendation system by more than 10 times, thereby accelerating revenue growth. This requires simultaneous investment in inference and training computing power. Larger, more expensive RecSys models are already driving advertisers to pay higher prices while maintaining strong ad returns.

Despite all this, still holding painfully but staying optimistic.