As we approach the end of the year, it’s worth reflecting on the numerous breakthroughs and changes over the past year, particularly in generative AI. AI-related startups have received a staggering $68.7 billion in funding since the beginning of the year. To distill the world of Generative AI, here is a visualization of the market
Market Map for Generative AI
Foundational Models:
The release of ChatGPT by OpenAI, which wrapped GPT3 in a user-friendly interface, caused a sensation in the tech world. The platform now boasts 100 million weekly active users. This year has seen significant developments in AI models, from large proprietary models to small & device-compatible models. Language, text-to-image, and multimodal models have rapidly increased over the past year. The primary ones here include:
Open AI: ChatGPT accounts for 60% of the web traffic across GenAI tools. OpenAI currently has two million developers using the APIs, and over 92% of Fortune 500 companies use OpenAI technology.
Google: Closest competition to OpenAI with introducing the Gemini suite of multimodal models, which I wrote about here. In the poll with the article, 63% said that they think Gemini models were better than chatGPT. As of this week, Gemini Pro is now available on Google AI Studio for developers to build upon.
Microsoft: The primary backer of OpenAI has seen massive gains this year, with their stock up 50% YTD. After using GPT-4 for Github Co-pilot & Microsoft 365 Co-pilot, they announced Phi-2, a small language model trained on 2.7B parameters. This model is small enough to run on a laptop or mobile, likely in response to Gemini Nano by Google.
Amazon invested $4B in Anthropic and launched Q, a chatbot for its AWS customers, trained on 17 years of AWS knowledge. They also released Titan: a text-to-image model for developers to build image generators.
Anthropic, founded by Dario Amodei, former VP of research at OpenAI, launched Claude 2.1. This model’s claim to fame is the context window of 200,000 tokens, translating to around 150,000 words. Users can upload large amounts of documentation, such as “The Iliad,” which has 176,000 words to summarize and interact with the content. How about a conversation with Hector or Achilles on a cold December evening?
Meta faced initial difficulties with Galactica, their language model trained on a vast dataset of 48 million scientific articles, websites, textbooks, lecture notes, and encyclopedias. Earlier this year, they released Llama2, an open-source language model in 3 sizes: 7B, 13B, and 70B parameters. It became the most popular open-source model until Mixtral surpassed it.
Mistral, founded by ex-Meta and Google researchers, launched its latest model, Mixtral8x7B. Within seven months of founding, the company raised $415M in funding at a $2B valuation. This is all thanks to its open-source model, comparable to GPT-3.5 & Llama2 on key benchmarks despite its relatively smaller size. GPT 3.5 was trained on 175B parameters, compared to Mixtral, which comprises eight expert models with 7B parameters. Mixtral is small enough to run on a single computer and only requires 100 gigs of RAM.
Storage & Compute:
On the chip side, NVIDIA emerged as the winner! NVIDIA’s GPUs have been in high demand among all GenAI companies, so much so that ChatGPT had to pause subscriptions because they ran out of computing. New businesses have sprung up that let you lease these servers from other people for those small idle moments that they have. NVIDIA’s stock price has increased by 235% during this time, and it has become the most prominent investor in upcoming Gen AI startups, having invested in 20 deals this year. Startups need GPUs more than money!
Not to be left behind, AMD unveiled the Instinct accelerator and the Instinct accelerated processing unit (APU). According to AMD, the Instinct accelerator has 2.4 times more memory capacity than NVIDIA’s GPUs and has 1.3 times more raw compute power. AMD also updated ROCm, its open-source GPU programming platform, which competes with NVIDIA’s proprietary GPU programming platform, CUDA.
Intel CEO Pat Gelsinger also revealed an Intel Gaudi3 AI accelerator, which is set to arrive next year. Amazon, Google, and Microsoft are also working on their AI chip investments to avoid relying on NVIDIA.
Amazon introduced Trianium and Inferentia chips earlier this year. AWS is now the company’s primary cloud provider as part of its investment in Anthropic. Anthropic will build and deploy future generations of its foundation models on AWS Trainium and AWS Inferentia chips. Google has its TPUs, which Gemini, Google’s general AI model, was trained on and served. At last month’s Ignite conference, Microsoft unveiled the Maia 100 artificial intelligence chip, which could compete with Nvidia’s highly sought-after AI GPUs.
The tech giants have been in cloud wars for a while now. The GenAI revolution added a few more names to the cloud service providers list -
CoreWeave is a cloud provider offering flexible and cost-effective GPU-accelerated compute resources on demand. It provides faster infrastructure and supports diverse compute-intensive use cases, making it stand out from competitors. It has investors such as Fidelity, Jane Street, JP Morgan Asset Management, NVIDIA, and Zoom Ventures. In December 2023, it was valued at $7 billion, up from the Series B valuation of $2 billion in April. Although it generates over $600 million annually, it is still far behind giants like AWS and Google Cloud, which had $80 billion and $26 billion in revenue last year.
Crusoe is a super-impressive company that uses low-cost, eco-friendly energy resources to power its compute infrastructure. In a time when the noise of the Climate Impact of AI is very loud, this looks like a step in the right direction. The company has a valuation of $1.7 billion and a revenue of approximately $100 million. Crusoe recently secured a $200M debt facility from Upper90, an investment firm, which it plans to use to purchase the latest NVIDIA chips. The interesting part is that the GPUs will be used as collateral for the loan, assuming that the company can repay the debt within 3.5 years and generate revenue from these chips for at least seven years. Coreweave is another company that has a similar setup with its debt facility.
Lambda Labs is a company founded in 2012 that provides workstations, servers, laptops, and cloud services specifically designed for engineers and researchers who work on deep learning and AI development. Recently, Lambda Labs has expanded its services to include cloud services with access to NVIDIA GPUs for AI training and inference. Many reputable organizations such as Intel, Microsoft, Amazon Research, Kaiser Permanente, MIT, Stanford, Harvard, and the Department of Defense are among Lambda Labs’ customers.
ML Platform & Libraries :
What they do is best explained as an F1 driver driving a Ferrari versus regular people driving a Ferrari. These companies help enterprises and developers build and fine-tune LLMs for their use cases. Key companies here are -
Together AI: A full-stack platform and cloud service for developers at startups and enterprises to build open-source AI is a challenge to OpenAI when it comes to targeting developers. Founded by Vipul Ved Prakash, their cloud platform offers training, fine-tuning, and production of AI models at lower prices than major cloud vendors. They also provide consulting services for building customized AI models and are involved in open-source AI research projects. Their cloud infrastructure, which runs NVIDIA GPUs and networking across AI cloud partners like Crusoe Cloud and Vultr, is custom-designed for high-performance AI applications. With custom infrastructure, they offer significantly better economics on pre-training and inference workloads. Some leading AI startups, like Pika Labs, NexusFlow, Voyage AI, and Cartesia, are building a new class of models on Together Cloud. Rumored to be generating approximately $20 million, they recently announced a $102.5 million Series B lead by Kleiner.
Mosaic ML: Acquired by Databricks for $1.3 billion, MosaicML is the generative AI platform that empowers enterprises to train, tune, deploy, scale, and monitor their own AI using their data along with leading open-source models, Mosaic’s models, or their models. MosaicML enables developers to maintain complete control over the AI models they build, with model ownership and data privacy built into the platform’s design.
Hugging Face is a platform offering various data science hosting and development tools. It can be compared to GitHub, but specifically for machine learning. The platform provides a hub-like environment for AI code repositories, models, and datasets. It also offers web apps that demonstrate AI-powered applications. Additionally, it offers libraries for tasks like dataset processing and model evaluation. Hugging Face has an enterprise version of the hub that supports on-premises and software-as-a-service deployments. The Hugging Face community has over 10,000 customers and over 50,000 organizations on the platform. Furthermore, its model hub hosts more than 1 million repositories.1
To recap, it has been a thrilling year with numerous significant advancements. The sector received an investment of $68 billion, and it is hoped that this money will be utilized efficiently. Below is a summary of all the companies from the market map and some essential details. Though this list has many noteworthy exclusions, such as Scale AI, I trust that it provides a good overview of the broader ecosystem.
Private Companies from the Market Map
Public Companies from the Market Map
If you have made it this far, thank you for reading. This will be our last post for 2023. We will start 2024 afresh with some key predictions for the year. Until then, Happy Holidays!
https://techcrunch.com/2023/08/24/hugging-face-raises-235m-from-investors-including-salesforce-and-nvidia/
Sources for Private company valuations, funding, and revenue
https://www.nytimes.com/2023/12/10/technology/mistral-ai-funding.html
https://www.businesswire.com/news/home/20230323005299/en/Personalized
Superintelligence-Platform-Character.AI-Secures-150M-in-Series-A-Funding-Led-by-Andreessen-Horowitz
Other Market Maps I found useful as I built my own -
https://www.sequoiacap.com/article/generative-ai-act-two/
https://www.bvp.com/atlas/state-of-the-cloud-2023#The-dawn-of-the-AI-era