Jeffrey Erickson | Content Strategist | January 12, 2024
When OpenAI released ChatGPT, its large language model (LLM), to the public in November 2022, it amassed 100 million users in just two months. That easily made ChatGPT one of the fastest growing consumer apps ever.
Investors took notice.
Since that time, members of Forbes’s top 50 AI companies list collectively have raised more than $27.2 billion. Some of those firms are under a year old with fewer than 20 employees. For AI startups with intriguing ideas, the cash is flowing.
Money raised, however, doesn’t guarantee success. AI startups take on unique challenges that call for more than the usual grit, market timing, and growth management. Training the LLMs that power services like ChatGPT or Midjourney’s AI image generator is one of the most computationally intensive tasks mankind has conceived. Investment firms say most of the capital that AI startups raise goes straight to computing resources.
Beyond that, AI startups are responsible for the security and privacy of sensitive information lurking in the mountains of training data their models ingest, even as they scramble to compete with incumbent giants that are also moving quickly to capture market share.
Microsoft founder Bill Gates calls AI the most significant development in computing since the graphical user interface, which launched Apple’s Macintosh and every popular operating system and application since. So it’s understandable that entrepreneurs want a piece of this action. Let’s look at factors that AI startups should be aware of as they enter the fray.
Generative AI startups come in three flavors: those that build LLM platforms, such as OpenAI or Cohere; those that offer new tools for building and training LLMs, such as MosaicML; and those that take open source LLMs and train them to solve specific business problems—an example is Tome, which applies AI to improve business presentations.
All AI startups are working in the afterglow of the likes of ChatGPT, Google, and other firms that have used powerful computing architectures called neural networks and machine learning algorithms to build friendly, natural language interfaces that can generate human-like text, visual content, and computer code and perform many other tasks.
From well-funded wunderkinds to bootstrapping upstarts, these startups face stumbling blocks unique to purveyors of AI-based services. The 11 challenges listed below provide a good sense of the possible barriers that await.
AI startups take on security and privacy responsibilities that go beyond standard corporate data protection efforts. Many security measures will be familiar, such as employing a zero-trust model and monitoring networks for malicious activity that set off automated responses and alerts. But there are also new challenges. For example, AI models potentially could leak details from the data used to train them. These data sets can be hundreds of gigabytes, or even terabytes, in size, pulled from a range of sources. They may contain sensitive data, including names, addresses, and other personally identifiable information. Might a model trained using this data reflect private details in its output?
It’s important for a startup to know what data is in its training sets and have a plan to minimize the risks involved with sensitive or regulated information. These companies need to satisfy investors that they have these concerns covered—and have a communications response plan in case something goes awry.
AI companies train and deploy large language models (LLMs) with extensive data sets and billions of parameters for all sorts of use cases, including natural language processing (NLP) and image creation. They also develop AI models for computer vision, forecasting and prediction, anomaly detection, and much more. LLMs in particular require incredibly large volumes of data to produce accurate and consistent outputs.
If you’re an AI startup, data management is at the core of your business.
A key challenge, however, is to find the right data sets for your AI training needs and upload them into a massive data warehouse or data lakehouse. Then data must securely flow through neural networks and machine learning algorithms using superclusters of graphics processing unit (GPU) servers—if you can find them.
A GPU is a chip with many more cores than a central processing unit (CPU). This design, as seen in Nvidia’s CUDA, short for compute unified device architecture, powers the massive parallelism required for tasks such as training AI.
An assertion we see over and over in TV shows, movies, and the popular media that AI will destroy the world. One counterargument: “Where will the evil AI get the GPUs?”
To run the neural networks on which AI models are built, GPUs split up computational work. The system then runs queries through a bunch of GPUs in parallel. That takes the load off the computer’s CPU and allows the network to crunch through complex calculations very quickly. Training and running AI models requires so much computing power that the world’s chip manufacturers and cloud providers are having a hard time keeping up with demand. Be aware you might need to get in line to buy chips or convince a cloud provider that your AI startup is worthy of those precious GPUs.
It’s safe to say most AI startups will build their companies around an LLM developed by another firm because, in most cases, customizing an AI model from the likes of OpenAI or Cohere is more efficient than designing, building, and training one from scratch.
There are two common approaches to customizing an LLM for a particular industry or use case: fine-tuning and retrieval-augmented generation (RAG). You fine-tune an AI system’s outputs by training it on large amounts of data specific to your cause and instructing the AI to give that information more weight in its responses. The other option, RAG, involves embedding highly relevant documents in a database that the AI will use to give context to the written or verbal prompts it receives. With RAG, those documents allow the AI to add relevant technical details to its output, and even cite where it got the information. For example, a healthcare startup might embed documents or articles that help its LLM better understand the intent of prompts from medical professionals and then provide output language related to their specializations.
Each method has benefits and drawbacks in terms of speed, quality, and cost. The approach to LLM customization is an important decision for any AI startup that hopes to deliver an industry- or use-case-specific service.
For companies in fast-moving startup mode, it’s hard to say no to ready-made cloud infrastructure. All hyperscale cloud providers offer what’s needed to train or customize large language models, including clusters of compute instances connected by a high-bandwidth network and a high-performance file system. And because these services are consumption-based, it’s often less expensive, and typically much faster, than setting up on-premises infrastructure.
Because these systems are consumption-based, speed and efficiency must be weighed against cost. An AI startup can keep its spending down by running an LLM that does what it needs to do with the least complex algorithms and least data possible. Once that budget calculation is made, choose a cloud infrastructure that handles your model efficiently. For example, running on bare metal servers avoids the overhead of virtualized instances and offers better performance. This becomes even more significant when it comes to clustered workloads common to LLMs.
Remember, the faster your job runs, the less you pay.
It can take many gigawatt hours of energy to train an LLM. For reference, a gigawatt could power as many as 874,000 homes for one years, according to investment firm The Carbon Collective. A startup seeking VC funding to deliver LLM-based services must prove it’s using its money wisely. For example, not all AI tasks need the same level of model sophistication or computational power. A growing collection of LLMs from companies such as OpenAI, Cohere, Anthropic, and others offer different variants and sizes. Be ready to explain why your choice suits your needs and your budget.
Once you’ve picked your model and data sets, carefully select an infrastructure with efficient parallel processing and dynamic scaling to avoid paying for compute resources you’re not using. Be prepared to show investors that your choices strike a balance between performance and affordability.
There are three primary techniques for scaling LLMs to increase the quality and/or speed of the LLM’s outputs: Increase the amount of training data, use a larger and more complex model, or add compute capacity.
A larger model increases the number of layers and parameters in the neural network architecture, giving it a higher capacity to learn and represent complex patterns in the data. As a result, your LLM will give more detailed and nuanced answers. By adding more gigabytes of training data, your AI startup can offer more accurate or complete responses. In both cases, you’ll also need to scale up expensive compute resources to maintain model performance.
This isn’t a challenge specific to artificial intelligence. Business analysts have been sweating the quality of the data they use for decades. AI startups need to tap into the expertise of data scientists and subject matter experts to remove redundant information, irrelevant content, and other “noise” from the data sets used to train algorithms and feed LLMs.
“Garbage in, garbage out” is an adage that should resonate with AI startups.
It’s important for AI startups to establish both quantitative and qualitative measures for success. Quantitative measures include ROI on technology investments and technical key performance indicators (KPIs), such as mean squared error (MSE), which identifies outlier results.
Beyond that, an AI startup should be able to measure qualitative results, such as how well an AI model performs on new or not previously seen data, how relevant results are for the target audience, and how comprehensive results are in context of the field being discussed.
There are a range of approaches for funding an AI startup. You can follow the example of LLMs such as Midjourney and Surge AI, which have grown their customer bases gradually without taking investment money. If your AI startup can’t wait for bootstrap growth, there are angel investors, accelerators, and incubators all looking for AI founders with sharp minds and good ideas. The benefit of incubators and accelerators is that they provide relationships, access to market opportunities, business advice, and even technology platforms for building an AI service.
Cutting-edge sales and marketing platforms are employing AI at every stage of the customer journey, and any AI startup looking to grow market share will want to employ AI to help. How? AI can use detailed data, including real-time geolocation data for mapping and tracking movement, to create product or service offerings personalized for potential customers. AI assistants can then generate upselling and cross-selling opportunities or nudge shoppers to complete transactions once items are in their carts. These tactics are proven to increase conversion rates and please investors eager for a startup to grow sales.
Post-sale, AI-enabled services can handle queries, understanding context and offering suggestions while also sharing concrete details about scheduling or delivery times and directing more complex questions to human agents. Seeing how these AI-backed services work can help you benchmark the offerings of your own AI startup.
If you’re building an AI-based business, consider Oracle Cloud Infrastructure (OCI), which provides a robust infrastructure for training and serving models at scale. Through its partnership with NVIDIA, Oracle can provide customers with superclusters, which are powered by the latest GPUs and connected with an ultralow-latency RDMA over Converged Ethernet (RoCE) network. This architecture provides a high performance, cost-effective method for training generative AI models at scale. Many AI startups, including Adept and MosaicML, are building their products directly on OCI.
Oracle makes it easy to get started with OCI services, including select Always Free cloud services. Startups can learn with developer sandboxes or packaged deployments of popular software, such as by deploying a Kubernetes cluster.
To help startups decide, Oracle provides exploratory tools, including cost calculators, third-party analyst reviews, and detailed comparisons between OCI and other cloud platforms.
Artificial intelligence has been in our lives for more than a decade, working in the background, monitoring for fraud amid millions of bank transactions, stepping in to handle frontline customer service interactions, and making lightning-fast decisions to speed up overnight shipping logistics. Now, with the latest generation of LLMs, the subtle, powerful, uncanny abilities of AI get the user interface they deserve: the natural spoken or written word.
As a result, LLMs have captured the popular imagination with image creation, written text and translation, and even code generation. Although challenges abound, now is the time for AI startups to find investors, serve new customers, and scale up like it’s 1999.
Establishing an AI center of excellence before organization-specific training commences makes for a higher likelihood of success. Our ebook explains why and offers tips on building an effective CoE.
What are common challenges for AI startups?
AI startups face challenges in picking the right LLM to train, finding the right training data, and assembling the immense computing power needed to support their neural networks. There are also issues of data privacy, data security, and shifting regulations to contend with.
What kinds of services do AI startups offer?
AI startups are popping up in every business sector, including ones as varied as healthcare, manufacturing, and national defense. Some startups are offering products to consumers while others are building tools used by other AI companies to build and train their models.
How do AI startups find funding?
Startups looking for investors can do a quick search for angel investors, who are on the lookout for opportunities. Other options are technology incubators or accelerators, which can provide guidance and technology assistance for startup founders.