All news

Baidu's Kunlun chip turns AI's largest buyers into chip designers

An open metal mould casting one cobalt square tile beside a short row of plain graphite tiles

Baidu unveiled an AI chip rated at 260 trillion operations a second. The largest buyers of AI hardware now design it, narrowing the market for merchant chips.

Baidu now designs the processors that run its own neural networks. Kunlun, unveiled at the company's developer conference in Beijing on 4 July, comes in one version for training models and one for running them, and Baidu rates it at 260 trillion operations a second on about 100 watts, roughly 30 times the speed of the programmable accelerators it has built on since 2011.

Kunlun matters less as hardware than as a shift in who buys AI silicon. The internet companies that run the largest networks were meant to be the anchor customers of merchant chip makers, the firms that sell processors to anyone. One by one they are becoming their own suppliers, and that removes the most valuable slice of the market from the plans of Nvidia and of the start-ups now raising money to compete with it.

The pattern is no longer one company's experiment. Google disclosed its tensor processor in 2016 and since February has rented it to outside developers at $6.50 an hour for a board of four chips rated at 180 teraflops. In April Alibaba said it is developing a neural network chip of its own and bought a domestic chip designer to speed the work. With Baidu, three of the largest operators of machine learning in two countries now design their own silicon. Their motive shows in the supplier's results: Nvidia's data center revenue reached $701 million in the quarter to April, up 71% from a year earlier, a margin the biggest buyers would rather keep.

Chinese buyers have a second reason. In April Washington barred American suppliers from selling to ZTE, cutting one of China's largest telecom equipment makers off from the chips inside its products. The ban was lifted only on 13 July, after ZTE paid a $1 billion penalty and placed another $400 million in escrow. Any Chinese company running its services on American accelerators now prices supply as a risk, and an in-house chip is the hedge.

For founders the consequence runs in two directions. Chip start-ups need customers that will never design their own silicon, such as carmakers, device makers, hospitals and factories. Application companies should keep their models portable, because the cheapest hardware to run a network on will keep changing. Fanaura favours that second group, the AI-first companies we invest in, whose margins improve whichever chip wins.

Designing a processor is the easy part, and that is the strongest case against reading too much into Kunlun. Baidu has set no release date, has not said whether it will sell the chip to anyone else, and still works with Nvidia to tune its own deep learning framework for Volta GPUs. Developers stay where the tools are. The deciding signal is whether Kunlun runs production work in Baidu's cloud at a published hourly price within a year, as Google's processors already do, a test we will apply across our notes on the chip market.

When the largest customer can build the product, the supplier's market becomes whatever that customer leaves behind.

For founders

Building AI products that run on whichever chip is cheapest, or chips for buyers that will never design their own? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.