Regarding the information gap on whether computing power is sufficient, looking at the actions of top AI companies speaks louder. The conclusion is that the current computing resources need structural optimization rather than being far from enough or having no upper limit. 1. According to Morgan Stanley's estimates, Meta has 3GW of computing power. By 2026 and 2027, Meta will add 4.5GW and 4GW, respectively. This means their total computing power will reach 7.5GW in 2026 and 11.5GW in 2027. Meta's inference computing efficiency is about 65%, so selling computing power is both a proactive choice and a necessity. 2. OpenAI and Anthropic currently have 1.9GW and 1.4GW of computing power, respectively. By the end of 2026, their reserves will reach 4–6GW and 3–4GW. 3. SpaceX's computing power lease agreement with Google is priced at 4x the industry standard; their agreement with Anthropic is priced at 3x. While the contract is nominally for three years, it's effectively a three-month auto-renewing agreement. This highlights a few issues: - Spot computing power leases come with a premium, indicating a short-term shortage. - Anthropic and Google opted for short-term agreements instead of long-term ones, suggesting this computing power is for flexible needs rather than long-term demand. - The shortage of computing power is structural. Companies like Meta and SpaceX are hoarding resources, having overestimated the computing power their products would require. Now, they're renting it out as a cost optimization strategy. 4. In China, the two companies with the most computing power are ByteDance and Alibaba, whose resources are almost fully utilized. Current estimates put their computing power at 2–4GW, expected to increase to 5GW by the end of this year. This is on par with OpenAI and Anthropic, but they don't plan aggressive expansions by 2026, possibly due to domestic restrictions. 5. DeepSeek's computing power is at 0.1GW. Their API pricing only differentiates between peak and off-peak times, yet they can serve a global audience. Is there no shortage of computing power? Not sure how they’re doing it, but I lean toward the idea that there’s still massive room for optimization in current computing resources. 6. Both DeepSeek and OpenAI are aggressively pushing for computing cost optimization. In early July, OpenAI revealed it could cut AI model inference costs by more than half (over 50%). DeepSeek has reduced inference caching costs by 90%, and other open-source models are following suit. 7. The ones truly short on computing power are OpenAI, Anthropic, and Gemini. Meanwhile, lagging models like Meta and SpaceX have shifted to renting out their computing power. 8. Data from OpenRouter shows that token consumption growth has significantly slowed in recent weeks, and there’s a shift from high-cost AI to low-cost AI.