Alibaba Cloud lowers LLM costs to record lows in China, causing a price drop in the market.
In the realm of artificial intelligence, a price war has erupted this year, shaking up the large language model (LLM) market. The catalyst for this change was the release of the DeepSeek V2 model on May 6, boasting an impressive 236 billion parameters.
DeepSeek V2 offers API pricing that is remarkably affordable, with computation costing RMB 1 (USD 0.14) per million tokens and inference costing RMB 2 (USD 0.28) per million tokens. This competitive pricing is part of a larger trend in the industry, as companies strive for training efficiency optimization and economies of scale.
Baidu, a key player in the market, reduced the price for its Wenxin model in 2023, lowering the inference cost to just 1% of its original cost. This reduction was accompanied by a change in token calculation, matching the number of Chinese characters, which effectively lowered prices by 20%.
Similarly, Alibaba Cloud reduced the price for its Qwen-Long LLM by 97%, from RMB 20 (USD 2.7) to RMB 0.5 (USD 0.06) per million tokens, on May 21. This move undercuts ByteDance's Doubao model, which is priced at RMB 0.8 (USD 0.11) per million tokens.
Microsoft joined the fray with the release of the 3.8 billion parameter model Phi-3 Mini, claiming performance comparable to GPT-3.5. Phi-3 Mini is capable of running smoothly on Apple's A16 chip.
However, despite these price reductions, GPT-4, with over 20 times the parameters, dwarfs Llama-3's performance.
On the other hand, smaller models could be leveraged for strategies in areas like data management and efficiency optimization. For downstream consumers, these models may offer better cost and budgeting benefits.
Notably, Lee Kai Fu, CEO of 01.AI, announced on May 21 that his company would not partake in the ongoing price war, retaining its current API pricing for its latest Yi-Large model.
Aside from the disruption in pricing, there is a renewed focus on smaller, cost-effective models. This shift is not just about price, as major companies tend to position large models as the "storefront," with the real intention being to sell complementary cloud services.
In response to Alibaba Cloud's announcement, Baidu made its Ernie Speed and Ernie Lite models free to use. This move demonstrates the competitive nature of the market, as companies vie for market share and consumer loyalty.
Meanwhile, Meta released the Llama-3 open-source model with 70 billion parameters on April 22. Despite its impressive size, it pales in comparison to the performance of GPT-4.
Lastly, High-Flyer, with its own computing cluster, can generate revenue of up to USD 35.4 per hour per server when the utilization rate of its computing service is at its peak, achieving a gross profit margin of over 70%.
This price war in the LLM market promises to reshape the landscape of AI, making it more accessible and affordable for a wider range of consumers and businesses.
Read also:
- Solar energy company, Imperium, alongside QORAY Mobility & Energies Solar Business, bolsters Nigeria's environmental future by producing superior solar panels domestically and offering flexible payment options.
- AI Inspection Company, Zeitview, Secures $60 Million Funding for Expansion
- Future of Payments: If the U.S. regulates stablecoins through the GENIUS Act, according to Matt Hougan
- Ongoing trade friction as the American administration levies fresh import taxes on goods arriving from China
