news

Author | Wang Wang, Editor | Manman Zhou

a16zAs an investment institution that is extremely optimistic about technology, a16z has made outstanding performances in every wave of technology. In addition to its outstanding investment projects, a16z is also creative in marketing and operations. Both its partners and invested companies are often on the hot search list.

Recently, a16z came up with a new idea and launched an "Oxygen Plan" to capture more AI startups on the market.

Computing power is a bottleneck

a16z wants to be the "oxygen generator" of AI companies

Computing power, algorithms, and data are the three cornerstones of artificial intelligence.

Since AI entered the era of large models, the demand for these three cornerstones has been increasing day by day. Since training models requires a large number of graphics cards, GPUs have become the hard currency in the AI ​​industry and were once in short supply.

According to The Information, a16z has stockpiled thousands of AI chips, including Nvidia's H100 GPUs, and rented them out to invested companies in exchange for equity. This is called the "Oxygen Plan" within the company, which means providing AI startup teams with the "oxygen" they need to survive. At present, a16z has invested hundreds of millions of dollars and will expand to a training cluster of more than 20,000 GPUs in the future, which is roughly the same size as the GPUs used by Musk to train the Grok large model.

a16z found that although the computing power shortage of AI companies has eased compared to last year, the market supply and demand is still very unbalanced. Compared with large companies that are generous with their spending, early startup teams find it difficult to obtain sufficient GPU resources. a16z believes that this will be an important bargaining chip to win the hearts of high-potential founders.

It turns out that a16z's Oxygen Plan does attract excellent entrepreneurial projects.Big Model CompanyLuma AIIn January of this year, it received a $43 million Series B investment from a16z. At that time, other institutions offered higher valuations, but a16z offered to rent GPUs at a low price and eventually won the negotiations.

Luma AI's co-founder and CEO once said that for AI infrastructure, computing is everything. Six months after completing the financing, Luma AI launched its first large-scale model of Wensheng video, Dream Machine, which is said to be comparable to Sora in online evaluation.

Image source: SiliconANGLE

Computing power for equity

This generation of investors loves to "sell shovels"

a16z is not the first investment firm to use GPU computing power in exchange for equity in invested companies.

Earlier, Silicon Valley investors Nat Friedman and Daniel Gross purchased 2,512 H100s, and institutions such as Index Venture and Convinction also purchased GPU computing resources to provide to start-ups.

VC institutions are like this, and cloud vendors are even closer to investing in AI companies.OpenAI、Mistral AI(OpenAI's competitor in Europe), including Alibaba's acquisition of MiniMax,Dark Side of the Moon, Zero One Everything,Zhipu AI, Baichuan Intelligence, and five other large-scale model unicorns. All of these investments emphasize "computing power to support cooperation." It is rumored that nearly half of the $800 million Alibaba invested in Dark Side of the Moon was cloud credits rather than cash.

Computing resources are extremely important for AI companies | Source: Gcore

For market-oriented institutions, providing scarce computing resources is undoubtedly a killer weapon in the current competition for excellent projects; for cloud vendors, this is a simple way to establish ecological synergy and increase their own revenue. AI companies are at the forefront, with volume models, volume applications, and volumePMF, investors sold shovels in the back and also rolled out a different business model.

Whoever controls the GPU controls the world?

Some people are not optimistic about the hoarding of GPUs. "Now that we have so many "shovels", what will we do with them in four years?" one netizen asked sharply.

Some investors hold the same view, such as Sequoia Capital. In this wave of AI investment boom, Sequoia has adopted a cautious attitude. Although it has made many investments in the past two years, the amount is not large, and it prefers early projects. On the topic of GPUs, Sequoia partner David Cahn believes that September 2023 will be the peak of GPU shortage, and the shortage has begun to ease. With the rapid development of technology, this generation of GPUs will depreciate significantly. The B100 launched by Nvidia in March this year has a performance 2.5 times that of the H100. In the future, the output will increase, and the H100 in a16z's hands will only become less and less valuable.

Convinction, the investment institution mentioned above, has slowed down its purchasing pace this year and sold some servers; Alibaba is also reportedly planning to quickly profit from the batch of GPUs it has before the new generation of AI chips are launched on the market.

In the AI ​​boom, will whoever has the GPU win the world?

Looking back, from the early years“Rent for equity”, to the era of mobile Internet“Traffic for equity”, and then to the AI ​​era“Computing power for equity”——The tricks have been upgraded, but the rules of the game remain the same, and in the end the "landlords" are the ones who reap the benefits.

According to Goldman Sachs analysts, Nvidia's earnings conference at the end of August will once again exceed revenue expectations. Things change, and only Huang is laughing.

Interaction at the end of the article:

What do you think of the "Oxygen Project" launched by a16z?

Tell us what you think in the comments section~