Tan Dai, President ​of Volcano Engine: Judging ⁢from ⁢the call volume, the⁤ cost ⁤of ‍large models is ⁣no longer an ⁤obstacle to innovation

Since‌ ByteDance released ⁤its self-developed ⁣Doubao‍ big⁣ model⁤ in May this year and‍ reduced the API call price to 0.0008 yuan per thousand tokens, which was 99.3% lower than the industry price‍ at the⁣ time. Four months later, ⁤the average daily number of Doubao​ calls has increased more than 10 times.

At the Shenzhen stop of Volcano Engine’s AI Innovation Tour today, Volcano Engine ​President Tan Dai announced that as of September this year, the ‌average daily usage of the Doubao model‍ exceeded 1.3 trillion tokens (the average daily call volume in ‌May this year was 120 billion times), with an ‌average of‌ 50 million images generated per day and 850,000 hours of voice processed per ⁣day.

“The second half of the ⁢year is the first year of the ​entire AI application, and this figure further confirms this view.” Tan Dai⁢ said.

Tan Dai, President of Volcano Engine

“When ‍the ⁤price drops by one‌ tenth, the⁣ volume may ⁤increase tenfold.” As for whether the low-priced‌ large model is sustainable, Tan Dai said in an interview‌ with ⁢the Science ​and Technology ⁢Innovation Board Daily and other media, “Their ⁣current focus is on application coverage ⁣rather than revenue. To unlock new scenarios,‌ we need stronger model capabilities, which we think is more valuable.”

In Tan Dai’s‍ view, only with‌ large ⁤usage can a good model be polished and the unit cost of⁣ model inference be greatly reduced; and ​when the price ‍of large models changes from cents​ to cents,‍ it can also help companies accelerate ‍business innovation at ‌a lower ⁤cost.

However, Tan Dai also believes that the business​ facing the B-end market ⁣must first be⁣ sustainable, and cannot rely on advertising for profit like‌ the 2C business.‌ He does not agree with the outside world’s view that the business⁤ of ​large model​ manufacturers has negative gross profit. ‍”2B products need to achieve ⁣positive gross profit, and we ‍have⁤ the ability and confidence to do​ so.”

Volcano‌ Engine’s pricing⁤ strategy for large models has attracted other ⁢manufacturers to follow suit.

Taking Alibaba Cloud as an example, after the‌ price of⁢ the⁤ GPT-4 ‍main model ​Qwen-Long was reduced by 97% in ⁢May, the prices of the three Tongyi Qianwen main models ‌on the ⁣Alibaba Cloud⁤ Bailian platform ⁢were ⁢also significantly reduced,‍ among which the price of ⁣Qwen-Turbo dropped by 85%.

“Alibaba’s⁣ price cut is good this time. The first time, they​ didn’t lower it to‌ the bottom, but now they are the ​same as us.” Tan Dai said that cost used to be ‍an ‍obstacle to innovation,‌ but now after the price cut, it is no longer an obstacle in terms of the number of calls. “The next thing ⁣to do is to⁢ improve the quality and performance based on this price. Quality ⁤means making⁤ the model more‍ powerful and more diverse.”

Currently, the Doubao large model family has covered large ‌language models such as the Doubao general model, ⁤role-playing model, and vectorized model, large visual model products such as the text-to-graph model and the graph-to-graph model, as well⁤ as large speech models⁣ such as the speech‌ recognition model and the speech synthesis model.

Today⁣ (September‌ 24), Volcano Engine announced that the Doubao big model has ‌added a video generation model, and also released the Doubao music model and simultaneous interpretation⁣ model, ‍achieving full modality coverage of language, voice, image, video,‌ etc.

Among them,⁣ the newly ⁣released Doubao video generation model⁣ includes⁤ two large model products, PixelDance ⁣and Seaweed, and has opened invitation testing for the enterprise market. It is​ reported⁢ that the Doubao video generation model has been tested on a small scale in​ the internal beta version of Jimeng AI, and will be ​uploaded to the Volcano Ark platform after ‌this year’s National Day for reservation ‍and use.

The pricing of the Doubao video generation‍ model has not yet been determined. ‌Tan Dai said that the application scenarios of video models and language models are ‍different, and the pricing logic is also different.⁢ For example, it is necessary to comprehensively ⁢consider new experience, old experience, and migration ‌costs. ‍Whether they can be widely‌ used in​ the end depends on whether the productivity ROI is much improved compared to before.

Volcano Engine has emphasized full-stack optimization of software and hardware since the first day of ‌cloud computing, but this does not mean that it has to ⁤do everything by itself.⁢ Hardware R&D refers more to ⁣computing, ⁣storage, and⁤ networking, and it is necessary to do​ end-to-end⁤ optimization and good combination.

“For‍ example, we make the DPU and video codec ​chips⁤ ourselves, and cooperate with friendly companies ⁣on CPU, ⁢GPU, etc., and use‍ our engineering capabilities to ⁢carry out hybrid cascade​ scheduling to improve performance and cost.”‍ Tan Dai said⁤ that​ one of the reasons why Doubao Token can achieve sustainable low prices is​ that they have made a lot of optimizations between software and hardware.

Entering the ⁤AI ​​era,⁣ Tan Dai‍ noticed that the changes ⁣in the B-end market are ‌that, on the one ‌hand, corporate⁤ demand has become⁢ slower, and‌ the core demand is to reduce costs and increase efficiency; on⁢ the other hand, AI applications have changed from top-down planning to bottom-up driven​ innovation.

Therefore, for large model manufacturers, the challenge‌ becomes moving from volume pricing to volume ⁣performance,‌ volume better model capabilities and services.

Tan Dai said, “The application⁢ cost of‍ large models ⁢has been well solved, and price is‍ no longer a bottleneck. In the future, we need to maintain cost-effectiveness ⁤and further improve our capabilities.”