When you talk about investments in artificial intelligence startups versus machine learning startups, it’s important to distinguish “AI” from “machine learning.” Those phrases are often used interchangeably, but they carry a slightly different meaning.
Machine learning, or ML, is a method of training AI models so that they can learn to make decisions. Put another way, ML involves training models to solve specific tasks by learning from data and making predictions. AI, on the other hand, is the broader concept for systems that mimic human cognition.
So ML is a subfield of AI but not the same thing.
Lonne Jaffe, managing director at Insight Partners, explains that Insight uses a “three-layer” framework to unpack the definition of an ML startup.
We’re widening our lens, looking for more investors to participate in TechCrunch+ surveys, where we poll top professionals about challenges in their industry.
If you’re an investor and would like to participate in future surveys, fill out this form.
At the first layer, he says, are core infrastructure companies — products with which a person builds an ML system. At the second layer are apps that seek to tackle a particular use case or workflow using ML. The third layer, meanwhile, comprises ML startups that manifest within an industry as an “actual player” in that industry — think startups that become a startup bank, even if the core of the startup is still ML talent.
According to this framework, examples of ML startups range from Weights & Biases, which provides tools to create and monitor AI models, to Iterative Health, a healthcare company that leverages an ML system designed to identify cancerous polyps from a colonoscopy.
The market for ML is quite large, with a report from Grand View Research estimating that it was worth $49.6 billion in 2022 and could grow at a CAGR of 33.5% by 2030. And it’s been building for some time: A 2021 survey by Dresner Advisory Services found that 59% of all large enterprises are deploying ML, with 50% of those organizations claiming to have 25 or more ML models in use today.
Why is this area growing so fast? 451 Research, the tech R&D group within S&P, posited in a recent report that the initial wave of ML adoption focused on making legacy systems and processes smarter — like business intelligence, customer support, sales and marketing and security. But now, as those applications mature, the attention has shifted to more niche, industry-specific and lucrative ML applications, particularly in finance, retail, manufacturing and healthcare.
Jerry Chen, a partner at Greylock, believes we’re just starting to see what the next generation of ML companies will be. “The cycle is going strong,” he told TechCrunch+. “I’m curious to see how incumbent companies and tech players enter, compete or partner with the startups. In particular, I think we will see some interesting go-to-market partnerships in the next few months.”
But what about the broader VC ecosystem? Are VCs in general optimistic about the future of ML?
To get a better sense, TechCrunch+ surveyed investors including Chen and Jaffe about the state of ML investing today. We touched on the health of the ML funding landscape, and whether the hype around ML, which several years ago was quite strong, is beginning to die down. We also asked investors what challenges stand in the way of ML tech adoption and what the next few months might look like in terms of market growth.
We spoke with:
- Lonne Jaffe, managing director, Insight Partners
- Jerry Chen, partner, Greylock
- Ashish Kakran, principal, Thomvest
- Janelle Teng, VP, Bessemer Venture Partners
(Editor’s note: The following responses have been edited for length and clarity.)
Lonne Jaffe, managing director, Insight Partners
How strong is the ML venture fundraising market today and how has it evolved thus far in 2023?
The release of ChatGPT five months ago sparked the fire of startup innovation around ML, along with a renewed fundraising dynamic. We’ve gone from systems of prediction — like classification or recommendation systems — to systems of creation. While funding has been flowing into generative ML systems, there has also been a lot of progress in more “traditional,” discriminative ML systems, like prediction or classification systems.
We’ve been particularly active recently in applied computer vision ML systems in healthcare, some of which may soon match or even exceed human physician performance across certain domains. For example, dental startup Overjet uses AI to analyze dental X-rays to help dentists decide if a tooth needs a filling or a crown, improving patient outcomes.
Unlearn uses ML to create synthetic digital twins — virtual patients created by combining AI predictions with existing patient profiles — to allow researchers to conduct clinical trials while needing significantly fewer patients in the control group to get the same quality trial outcomes.
Is the hype cycle in ML dying down or going strong? What might the next few quarters look like?
It’s been a wild year so far. From what we’re seeing, it’ll likely continue to accelerate for the rest of 2023. Before the release of ChatGPT and GPT-4, I think many people underestimated and underappreciated the progress that ML had been making over the last few years. Every week will bring new information, informing active public debates like large models versus small and open source models; fine-tuning ML models versus using vector similarity search and prompt enhancement; and startups versus incumbents capturing more value.
Engendering user trust, building partner ecosystems, creating efficient ML infrastructure, attracting and retaining high-caliber talent, and harnessing usage scale to gather feedback data will likely continue to be important ingredients in ML startup defensibility.
Are there technical barriers in the way of success (e.g., GPU capacity)? What about macroeconomic conditions?
The GPU shortage is a barrier, but each barrier can also present a new ML startup opportunity. For example, Run:AI is similar to VMware, but for GPUs instead of CPUs. If a company is performing ML model training or inference with specialty hardware like GPUs or similar chipsets, deploying Run:AI can help the system become faster and less costly, dramatically decreasing the number of GPUs it needs. Other challenges that also represent startup opportunities include privacy, security, intellectual property issues, explainability, understanding causality, hallucination and reliability.
The current macroeconomic conditions could actually accelerate the adoption of ML, similar to how the pandemic increased the adoption of collaboration technologies like Zoom. With the recent uptick of inflation, one way we can address supply and labor shortages is to intentionally slow down the economy using techniques such as substantially higher interest rates. But another, less harmful approach could be to use ML to dramatically improve productivity. In a way, ML could become the unsung hero of this inflationary crisis period.
What advice would you give to ML startup founders for growing their business and attracting investors?
In the end, creating extraordinary products with massive scale of usage will be critical. As ML systems often benefit from more use, we can expect value to accrue at the application layer for companies that truly understand their users’ pain points and needs, and build phenomenal product experiences around them.
Efficient runtime infrastructure, a strong focus on understanding the user experience and domain-specific data are also showing promise as sources of economic power. Investors will also look for efficiency metrics like high gross margins, strong gross retention, rapid expansion within customers, decreasing customer acquisition costs, shorter sales cycles and productive sales reps.
Gross retention, in particular, will be critical, because companies must be able to retain customers to stabilize their growth plans. In a difficult budget environment, high gross retention rates can be a strong signal that customers love your products and get real value from them.
Any other thoughts you’d like to add?
As adoption of ML increases, safety concerns will shift to the forefront. New technologies give rise to nuanced issues, opportunities and dangers that businesses and governments should focus on. As powerful ML tools become easier to use, we may see incumbents who are not native to ML retrofit existing products, especially as many ML capabilities are provided as an easy-to-use service from hyperscale vendors, including Big Tech companies.
Startups can create genuinely new capabilities, a “wow” experience that gets new users in the door, and build additional workflows around generative technology over time to retain users and increase stickiness.
Jerry Chen, partner, Greylock
How strong is the ML venture fundraising market today and how has it evolved thus far in 2023?
The ML funding landscape today is incredibly exciting and robust, but there is constant change. There’s still a lot of interest around foundation models, but I’m super excited about the evolving application infrastructure and data stack around building AI apps. Finally, we are seeing a new generation of AI-powered applications attacking different markets and verticals like CRM, security, legal and healthcare.
Is the hype cycle in ML dying down or going strong? What might the next few quarters look like?
The cycle is going strong. I think we’re just starting to see what the next generation of ML companies will be. I’m curious to see how incumbent companies and tech players enter, compete or partner with the startups.
In particular, I think we will see some interesting go-to-market partnerships in the next few months.
Are there technical barriers in the way of success (e.g., GPU capacity)? What about macroeconomic conditions?
I don’t think there are barriers that people can easily point to as much as areas of research that can change the market in the coming years, like distilling models or pruning models for different use cases.
What advice would you give to ML startup founders for growing their business and attracting investors?
ML can create a lot of value, so founders can be tempted to build a product that’s too broad. Instead, founders should find a great wedge. First, look for a problem and solve it — that’s the narrow part of the wedge — then earn the right to expand to the wider part of the wedge and address more problems.
Ashish Kakran, principal, Thomvest Ventures
How strong is the ML venture fundraising market today and how has it evolved thus far in 2023?
There’s a massive graveyard of enterprise ML projects that cost six to eight months and millions of dollars but failed to meet their stated objectives. It could have been because of a combination of issues around talent, data or compute, as companies outside of the best tech companies can be lacking in these areas.
I define ML startups as companies that solve these problems, enabling enterprise leaders to experiment, pick the best performing models, operationalize them and then monitor them in production for model decay. We believe in the cloud and ML infrastructure space, where often founders build differentiated products to solve issues that a large number of customers across multiple industries face.
Is the hype cycle in ML dying down or going strong? What might the next few quarters look like?
The ML funding landscape is certainly active after a general investing slowdown in 2022, which followed the valuation indiscipline of 2021. One of the catalysts for the current activity is the breakthrough achieved because of transformer models. They’ve certainly taken machines closer to the way humans think and respond to stimuli. When such a technology breakthrough happens, it forces enterprise leaders to reprioritize projects and budgets.
When the pandemic hit, many projects that normally would’ve taken years were completed in months. With generative AI, the technology that was expected in 15 to 20 years is now here, so companies both large and small are trying to figure out how to benefit from it. This presents an amazing opportunity for founders to solve unique, new challenges.
Are there technical barriers in the way of success (e.g., GPU capacity)? What about macroeconomic conditions?
We are witnessing a big change in the way humans interact with computing devices. Generative AI is the iPhone moment of our times, and this technology shift will impact every vertical and every function in the enterprise by making employees more productive, and, in some cases, eliminating the need to hire more.
Code that needed days or weeks to be produced in the past can now be generated in a matter of seconds. This type of disruption is being observed across use cases like content creation, legal document review and financial analysis.
The more important aspect to note is that, now, everyone is a developer — you just need to be good at writing prompts to get 80% of the solution to your problem. We will continue to see more investments in this space.
What advice would you give to ML startup founders for growing their business and attracting investors?
I’ve been a founding engineer and have closely seen startups that worked and others that didn’t. As engineers, we gravitate toward rapid prototyping when we have an idea, but many founders waste precious time building MVPs to solve problems that don’t matter.
You can save a ton of time by talking to 30 to 40 potential customers before writing the first line of code. My biggest suggestion is to work backward from customer problems instead of doing a “show-and-tell” of the amazing technology that you’ve built.
If you get the product-market fit right, then customers and investors are both likely to follow.
Janelle Teng, VP, Bessemer Venture Partners
How strong is the ML venture fundraising market today and how has it evolved thus far in 2023?
The ML funding landscape has been quite robust, in contrast to the overall VC activity slowdown. In particular, the model layer has received a lot of attention, with several LLM companies announcing mega-rounds within the last month. Some even raised massive amounts at the seed stage, which is striking.
But there has been a lot of innovation, and consequently funding, in other layers of the tech stack — from applications and infrastructure to tooling and orchestration, as well as startups building full-stack. We believe that enduring businesses can be built across all parts of the stack beyond just the model layer, and this movement is global.
Is the hype cycle in ML dying down or going strong? What might the next few quarters look like?
We are certainly in a bit of a hype cycle given the large amounts of VC funding and all the attention on ML from enterprises and consumers alike. But I’m reminded of Amara’s law, which states “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
Even with all the dollars already funneled into the category, ML, as a technological paradigm, is still early in its S-curve and adoption cycle; a lot remains unproven. While not all early ML startups will succeed, I firmly believe that the category will completely transform our lives and the economy.
For much of history, most ML advancements were only accessible to research communities. But today, the unprecedented democratization of access to ML advancements, such as through open source, is accelerating innovation and adoption exponentially. The incredible innovations we have witnessed so far are just scratching the surface of what’s to come, and ML will likely permeate many more aspects of our daily lives in the coming years.
Are there technical barriers in the way of success (e.g., GPU capacity)? What about macroeconomic conditions?
This current supercycle driven by ML has certainly caused a phase shift in demand for compute capacity, but these concerns should be resolved in due time by various factors, such as the velocity of chip innovation and entry of new market players. There are even startups taking creative approaches to tackle supply-side limitations, such as by improving efficiency and optimizing access to underutilized GPU resources.
To your question on macroeconomic conditions: The commercial interest and customer-readiness for ML solutions has remained resilient even in such challenging times as organizations recognize the immense power of ML to supercharge their businesses. In Morgan Stanley’s latest Q1 2023 CIO survey, AI and ML was the fourth-largest area to see enterprise IT spend increases, and the survey found that more CIOs are in the process of evaluating AI and ML technology this year.
I still see the most significant barriers in the way of success stemming on the execution side — in particular, around founders being able to bring academic theories or research into production as scalable and practical offerings. Many ML startups are truly operating at the cutting edge of what’s possible. While this is exciting, similar to frontier tech companies, many ML startups have not yet passed the threshold for technological or productization de-risking and will likely require a significant amount of specialized R&D resources and talent to execute on and scale their vision.
What advice would you give to ML startup founders for growing their business and attracting investors?
There’s so much noise in the market right now. I would advise you to be concise with your message around technical and product differentiation so that customers and investors — especially those who are nontechnical — have a clear understanding of your “secret sauce” or specific breakthrough that gives an explicit competitive advantage.
On top of this, make sure to communicate your business value proposition in a straightforward manner so that all stakeholders can come away with a tangible grasp of the relevance to their business. The most powerful conversations leave potential partners not just inspired by your vision but excited to deploy your ML innovation as soon as possible to unlock incremental value for their organizations.
Comment