Cohere's most recent funding was a Series C round in June 2023 where the company raised $270M at a valuation of $2.2B. The round was led by Inovia Capital and saw participation from investors including Nvidia, Oracle, Salesforce Ventures, DTCP, Mirae Asset, Schroders Capital, SentinelOne, Thomvest Ventures, and Index Ventures.
As of August, Tiger Global Management was in talks to sell a portion of its stake in Cohere at a valuation nearing $3B, marking a more than 40% increase from the last round in June. The proposed sale involves a 2.1% stake for approximately $63M, which is equivalent to Tiger's initial investment in Cohere. Post-sale, Tiger is expected to retain about a 5% stake in the company.
In September, Cohere engaged JPMorgan Chase & Co. and Goldman Sachs Group Inc. for assistance in raising further capital.
Cohere was co-founded in 2019 by Aidan Gomez, Ivan Zhang, and Nick Frosst, the latter of whom was among the initial hires at Google’s Toronto AI lab.
Before founding Cohere, Gomez co-authored the paper "Attention Is All You Need"—one of the "founding documents" of modern artificial intelligence—which introduced the Transformer architecture, a foundational technology for large language models.
Cohere, the company they started, is differentiating by focusing on enterprise applications of AI. Its offerings are unique from those of OpenAI in not confined to any specific cloud environment, allowing for deployment on public clouds like Google Cloud and Amazon Web Services, as well as customer-owned clouds and on-site servers.
Cohere is involved in custom model development, working directly with clients to leverage their unique data sets. While the company hasn't disclosed customer or revenue numbers, it has mentioned working with companies like Jasper and HyperWrite on copywriting tasks, LivePerson on conversational marketing, and news outlets on better summarization.
Cohere built Coral as a knowledge assistant for the enterprise that utilizes generative AI to assist in business operations. The model uses both internal and external data to provide responses specific to a given company or industry.
Unlike Command, Coral is tailored to cite internal data sources when providing information, an approach aimed at mitigating hallucinations in generated text and creating an "off-the-shelf" solution for businesses that want to launch LLM features internally.
It's designed to work with multiple types of data sources, such as customer relationship management systems and databases, and can be customized for the needs of different kinds of teams across the organization, from finance to support to sales to human resources.
Also, Coral is cloud-agnostic—allowing customers to deploy on any cloud from Oracle to Amazon to Google, as well as in a virtual private cloud or on-prem.
Command is a text generation model intended for a range of business applications across tasks like summarizing text, generating copy, and answering questions.
The model is updated weekly and is designed to be compatible with various cloud environments. By being fine-tuned on a company's internal data, it can customized to fit specific business language and requirements.
Additionally, Command can be integrated into other products, which allows companies to add text-generation capabilities without maintaining the AI model themselves.
Example use cases of Command include:
- Writing product descriptions for an ecommerce site, drafting emails to customers, providing example press releases for a marketing team
- Automatically categorizing documents, extracting information, or answering general questions about a dataset
Cohere makes money by offering its AI models to enterprises via API.
The company's Free Plan targets developers and small projects, providing rate-limited access to all API endpoints and community-based support through Discord, all at no cost.
For businesses requiring more advanced features and scalability, the Production Plan offers elevated ticket support, increased rate limits, and the option to train custom models. Pricing under this plan is pay-as-you-go, with costs per 1 million tokens set at $1.50 for input models and $2.00 for output models.
The Enterprise Plan is aimed at large corporations and offers custom pricing based on dedicated model instances, specialized support channels, and a variety of deployment options, requiring a consultation with the sales team for a tailored quote.
In addition to these tiered plans, Cohere has specific endpoint pricing that varies according to the type of service, such as text generation or classification:
- Generate and Summarize: $0.000002 per token
- Classify: $0.00005 per classification
- Embed: $0.0000001 per token
- Chat: $0.000002 per token
- Rerank: $0.001 per search unit
This multi-tier approach allows Cohere to capture revenue from a wide spectrum of clients, ranging from individual developers to large enterprises.
OpenAI hit $2B in annual recurring revenue at the end of 2023, up about 900% from $200M at the end of 2022.
OpenAI was founded in December 2015 as a non-profit dedicated to developing “safe” artificial intelligence. Its founding team included Sam Altman, Elon Musk, Greg Brockman, Jessica Livingston, and others.
OpenAI’s first products were released in 2016—Gym, their reinforcement learning research platform, and Universe, their platform for measuring the intelligence of artificial agents engaged in playing videogames and performing other tasks.
OpenAI’s flagship consumer product today is ChatGPT, which millions of people use everyday for tasks like code generation, research, Q&A, therapy, medical diagnoses, and creative writing.
OpenAI has about two dozen different products across AI-based text, images, and audio generation, including its GPT-3 and GPT-4 APIs, Whisper, DALL-E, and ChatGPT.
Anthropic is an AI research company started in 2021 by Dario Amodei (former VP of research at OpenAI), Daniela Amodei (former VP of Safety and Policy at OpenAI) and nine other former OpenAI employees, including the lead engineer on GPT-3, Tom Brown. Their early business customers include Notion, DuckDuckGo, and Quora.
DuckDuckGo uses Anthropic to provide “Instant Answers”—auto-generated answers to user queries.
Quora uses Anthropic for their Poe chatbot because it is more conversational and better at holding conversation than ChatGPT.
In March 2023, Anthropic launched its first product available to the public—the chatbot Claude, competitive with ChatGPT. Claude’s 100K token context window vs. the roughly 4K context window of ChatGPT makes it potentially useful for many use cases across the enterprise.
This Israeli-based startup has rolled out its own competitor to GPT-3, termed "Jurassic". Additionally, they have developed tools leveraging AI to aid users in writing.
Yoav Shoham, co-founder and a former director of the AI lab at Stanford University, emphasized AI21 Labs' commitment to revolutionizing reading and writing practices. While their first model paralleled GPT-3 in size, they have since introduced a smaller yet high-performing variant.
The company boasts a developer base of approximately 25,000 for Jurassic. Additionally, as of November 2022, they partnered with Amazon to offer Jurassic through its cloud AI service.
Today, AI21 Labs is at $273M raised, with a $1.4B valuation.
Launched in 2021 by Noam Shazeer, a former Google Brain researcher and one of the original creators of the transformer, Character.AI specializes in allowing users to create their own chatbots. Their chatbots can emulate various personas, including notable figures such as Joe Biden.
The company's primary intent is to empower users to build their own bots to solve a diverse range of use-cases through Character.AI vs. prescribing one way of interacting with the tool.
So far, Character.ai has raised $150M. Its investor roster includes tech pioneers like Paul Buchheit, Gmail's creator, and Nat Friedman, GitHub's former CEO.
Earlier this year, Google merged its DeepMind and Google Brain AI divisions in order to develop a multi-modal AI model to go after OpenAI and compete directly with GPT-4 and ChatGPT. The model is currently expected to be released toward the end of 2023.
Gemini is expected to have the capacity to ingest and output both images and text, giving it the ability to generate more complex end-products than a text-alone interface like ChatGPT.
One advantage of Google’s Gemini is that it can be trained on a massive dataset of consumer data from Google’s various products like Gmail, Google Sheets, and Google Calendar—data that OpenAI cannot access because it is not in the public domain.
Another massive advantage enjoyed by Google here will be their vast access to the most scarce resource in AI development—compute.
No company has Google’s access to compute, and their mastery of this resource means that according to estimates, they will be able to grow their pre-training FLOPs (floating point operations per second) to 5x that of GPT-4 by the end of 2023 and 20x by the end of 2024.
Meta has been a top player in the world of AI for years despite not having the outward reputation of a Google or OpenAI—software developed at Meta like Pytorch, Cicero, Segment Anything and RecD have become standard-issue in the field.
When Meta’s foundation model LLaMA leaked to the public in March, it immediately caused a stir in the AI development community—where previously models trained on so many tokens (1.4T in the case of LLaMa) had been the proprietary property of companies like OpenAI and Google, in this case, the model became “open source” for anyone to use and train themselves.
One particular area of interest for Cohere in expanding the functionality of their models is adding search and retrieval functionalities, with the objective of providing AI systems with more autonomous capabilities.
Long-term, Cohere plans to develop models that can perform more specific business tasks like scheduling meetings or filing expense reports.
That would allow Cohere to expand their audience and grow ARPU with existing customers while at the same time potentially threatening their B2B focus by getting them into competition with e.g. scheduling tools and expense management platforms that are building their own LLM-based software features.
Across all of these examples, what we’re seeing is that companies don’t want to be dependent on any single LLM provider.
One reason is that using different LLMs from different providers on the back-end gives companies more bargaining power when it comes to negotiating terms and prices with LLM providers.
Working with multiple LLM companies also means that in the event of an short-term outage or a long-term strategic shift, companies aren’t dependent on just that one provider and have a greater chance of keeping their product going in an uninterrupted manner.
This means that there's a world where companies like OpenAI, Anthropic and Cohere can all thrive in a multi-LLM world similar to the multi-cloud world we have where Google Cloud, Azure and AWS exist in stable equilibrium.
By entering the consumer space early rather than building for businesses, OpenAI has marked itself out as potentially competitive with any product building an AI product for consumers.
Since the launch of the GPT-3 API, there’s been a wave of companies building text-based AI products—see AI writing assistants like Jasper and Copy.ai. Jasper and Copy.ai built their businesses reselling OpenAI’s GPT-3 output at ~60% gross margin. Then OpenAI released ChatGPT, with which users can upload a batch of text and have it edited via a chat interface just as they could have within Jasper or Copy.ai.
Instead, they can fully focus on developing a product specifically responsive to the needs of businesses, which might mean higher customization, better integration capabilities, a stronger focus on scalability and reliability, white-labeling, or better data privacy controls.
This report is for information purposes only and is not to be used or considered as an offer or the solicitation of an offer to sell or to buy or subscribe for securities or other financial instruments. Nothing in this report constitutes investment, legal, accounting or tax advice or a representation that any investment or strategy is suitable or appropriate to your individual circumstances or otherwise constitutes a personal trade recommendation to you.
Information and opinions presented in the sections of the report were obtained or derived from sources Sacra believes are reliable, but Sacra makes no representation as to their accuracy or completeness. Past performance should not be taken as an indication or guarantee of future performance, and no representation or warranty, express or implied, is made regarding future performance. Information, opinions and estimates contained in this report reflect a determination at its original date of publication by Sacra and are subject to change without notice.
Sacra accepts no liability for loss arising from the use of the material presented in this report, except that this exclusion of liability does not apply to the extent that liability arises under specific statutes or regulations applicable to Sacra. Sacra may have issued, and may in the future issue, other reports that are inconsistent with, and reach different conclusions from, the information presented in this report. Those reports reflect different assumptions, views and analytical methods of the analysts who prepared them and Sacra is under no obligation to ensure that such other reports are brought to the attention of any recipient of this report.
All rights reserved. All material presented in this report, unless specifically indicated otherwise is under copyright to Sacra. Sacra reserves any and all intellectual property rights in the report. All trademarks, service marks and logos used in this report are trademarks or service marks or registered trademarks or service marks of Sacra. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any report is strictly prohibited. None of the material, nor its content, nor any copy of it, may be altered in any way, transmitted to, copied or distributed to any other party, without the prior express written permission of Sacra. Any unauthorized duplication, redistribution or disclosure of this report will result in prosecution.