文字轉影片的革命取決於驚人的運算能力：需要數百萬個 GPU

文字轉影片產生的前景激起了人們對人工智慧代幣的興趣，但廣泛採用將需要大量增加運算能力。據估計，支援 TikTok 和 YouTube 的創作者社群需要 72 萬塊高階 Nvidia H100 GPU，耗資 216 億美元，遠遠超過 Meta 和微軟等科技巨頭目前擁有的資源。這突顯了人工智慧生成視訊主流面臨的重大硬體挑戰和潛在限制。

Text-to-Video Revolution Hinges on Staggering Compute Power: Millions of GPUs Required

文字轉影片的革命取決於驚人的運算能力：需要數百萬個 GPU

The advent of text-to-video generation has ignited excitement within the crypto market, with AI tokens soaring following the unveiling of OpenAI's "Sora" demo. However, making this technology mainstream poses a formidable challenge, requiring an astronomical amount of compute power.

文字到影片生成的出現引發了加密貨幣市場的興奮，隨著 OpenAI 的「Sora」演示發布後，AI 代幣飆升。然而，使這項技術成為主流面臨著巨大的挑戰，需要大量的運算能力。

The Sheer Number: Hundreds of Thousands of GPUs Needed

數量龐大：需要數十萬個 GPU

A groundbreaking report from Factorial Funds estimates that a staggering 720,000 high-end Nvidia H100 GPUs would be necessary to support the content-creator communities on platforms like TikTok and YouTube. This number dwarfs the combined GPU arsenal of tech giants such as Microsoft, Meta, and Google.

Factorial Funds 的一份開創性報告估計，支援 TikTok 和 YouTube 等平台上的內容創作者社群需要 72 萬個高端 Nvidia H100 GPU。這個數字讓微軟、Meta 和谷歌等科技巨頭的 GPU 函式庫總和相形見絀。

Training vs. Inference: An Exponential Power Demand

訓練與推理：指數級的功率需求

Training text-to-video models like Sora requires colossal compute power. According to Factorial Funds, Sora requires up to 10,500 GPUs for a month's training and generates a mere 5 minutes of video per hour per GPU during inference. As adoption grows, inference will surpass training, demanding even more computational resources to produce new videos.

訓練像 Sora 這樣的文字到視訊模型需要巨大的計算能力。據 Factorial Funds 稱，Sora 一個月的訓練需要多達 10,500 個 GPU，並且在推理過程中每個 GPU 每小時僅生成 5 分鐘的影片。隨著採用率的增長，推理將超越訓練，需要更多的運算資源來產生新影片。

Nvidia's Dominance, but Not a Monopoly

Nvidia 佔據主導地位，但並非壟斷

Nvidia reigns supreme as the leader in AI-specific GPUs, but it's not the only player. Rival AMD offers competitive products, and its stock has witnessed a meteoric rise in recent years. Alternative options exist for outsourcing compute power to GPU farms, but these networks largely rely on gaming GPUs, significantly less potent than Nvidia's server-grade offerings.

Nvidia 是人工智慧專用 GPU 領域的領導者，但它並不是唯一的參與者。競爭對手AMD提供有競爭力的產品，近年來股價飛速上漲。有將運算能力外包給 GPU 農場的替代選擇，但這些網路很大程度上依賴遊戲 GPU，其效力明顯低於 Nvidia 的伺服器級產品。

The Hardware Hurdle: A Call for More Chips

硬體障礙：需要更多晶片

The promise of text-to-video generation hinges on a herculean hardware investment. Nvidia, with its annual production capacity of 550,000 H100 GPUs, falls short of meeting the projected demand. Combined, the twelve largest users of H100 GPUs possess 650,000 of the cards, with Meta and Microsoft collectively holding 300,000.

文字轉影片產生的前景取決於龐大的硬體投資。 Nvidia的H100 GPU年產能為55萬片，但未能滿足預期需求。 H100 GPU 的 12 大用戶加起來擁有 65 萬張卡，其中 Meta 和 Microsoft 總共有 30 萬張。

A Multi-Billion Dollar Endeavor

耗資數十億美元的努力

Acquiring the necessary number of H100 GPUs would incur a staggering cost of $21.6 billion, nearly matching the current market capitalization of AI tokens. Even if the financial hurdles could be overcome, the physical availability of these GPUs remains a significant constraint.

購買必要數量的 H100 GPU 將花費 216 億美元的驚人成本，幾乎與當前 AI 代幣的市值相當。即使可以克服財務障礙，這些 GPU 的實體可用性仍然是一個重大限制。

Conclusion: Mainstream Adoption Remains Elusive

結論：主流採用仍然難以捉摸

The allure of text-to-video generation is undeniable, but its widespread adoption faces a formidable challenge in the form of massive compute power requirements. While the premise holds promise for revolutionizing creative workflows, expecting mainstream adoption anytime soon is unrealistic. The road to unleashing the full potential of this technology requires a substantial increase in chip production and a concerted effort to address the immense hardware demands.

文字轉影片產生的吸引力是不可否認的，但其廣泛採用面臨著巨大的運算能力需求形式的巨大挑戰。雖然這個前提有望徹底改變創意工作流程，但期望很快就被主流採用是不切實際的。要充分發揮這項技術的潛力，需要大幅增加晶片產量，並共同努力滿足龐大的硬體需求。