Over the last year, we’ve looked at many Generative AI startups and the reality is that many of these startups are the exact same under the hood, with many teams trying to make similar Generative AI applications. Many incumbents with hordes of data are also gunning for the same market. Given the technical and complexity aura around AI, one would think that various aspects of AI technology itself will act as differentiators for these startups, but after looking at some of the startups, we find that AI offers no such play to most startups, unless they are building something that smart AI researchers at OpenAI or Google or Facebook find intriguing. But for most startups’ teams, with traditional SW DNA, it is wise to think about what kind of differentiation they can create from the beginning. Given AI is mostly going to be vertical specific, the founding team must think about everything from product capabilities, GTM, customer ROI, to pricing, business model, services etc. Many startups are thinking in lines of proprietary Data, latest AI Model, and cool UX. But we think the aspects of GTM, and product ROI will become much more critical in AI context as the play evolves. AI Models are for sure going to be interchangeable between opensource and paid models. On the data front - the more specific and the better the quality of data, the better the AI results, and hence the moat. The more specialized the target use case, the higher the chance of being the first mission-critical system in customer environment. But getting access to Data is iffy, sure startups may get some proprietary data, but unless startups find out a friction less GTM, getting that data will remain a dream. And AI computing infra was never a startup thing. So, what is left? At Service Ventures, we are sharing some of our general thoughts below and asking a few questions to which we ourselves have no answers for. But first, the market landscape.
Web of Strategic Investments in AI
Various players in the IT ecosystem – Semiconductor Suppliers, HW & Systems Integrator, Infrastructure Vendors, Cloud Service Providers (CSPs) and Enterprise SW Vendors, and even new Generative AI upstarts themselves are all placing their bets on the general AI trends as none of them wants to miss out on one of the biggest technology trends in IT. The limited quantity and quality of AI investments available have inflated the valuation with a web of entangled investments among players with different interests and influences in IT the ecosystem. Many generative AI startups such as Cohere, Anthropic etc. have raised astronomical amounts of capital from various industry players to create full stack offerings – from HW to use case specific applications and everything in between. As the Generative AI upstarts mature and become large players in the ecosystem, the conflict of interests of their backer will come to the forefront. Microsoft Corp.'s $10B investment in OpenAI is facing a regulatory probe in both the UK and EU for the validity of Microsoft’s minority shareholder's position and the degree of influence it enjoys on OpenAI. Should OpenAI be considered a part of Microsoft? Anthropic on the other hand has decided to partner with the two other cloud providers – Google and Amazon. As clear market leaders in AI have not emerged yet, many IT ecosystem players have built out a web of stakes instead of going for outright M&As.
Even the well-funded Generative AI startups have also realized these dynamics and are actively investing in other smaller startups that align with their industry vision. OpenAI itself has made several minority investments in AI startups via its OpenAI startup fund. Similar is the case for Databricks that has invested in many startups after raising multiple growth rounds itself.
The big three Hyperscalers — Microsoft, Amazon, and Google — are also investing aggressively in the space by offering cash and computing power to AI startups for their partnerships. Hyperscalers want to own all the three layers of the generative AI stack – HW+SW infrastructure, AI Foundational model, and Applications. OpenAI, Hugging Face, Anthropic and Cohere are the top foundational models in the market and a proxy war has emerged in the ecosystem via these models. Microsoft's big bet is OpenAI - has paid handsomely for exclusivity — with Microsoft's $10B investment propelling OpenAI to the most valuable generative AI provider in the world. Amazon and Alphabet (Google), have both chosen Anthropic as their leading horse in the generative AI race. Amazon and Google are also backers of open-source model directory specialist Hugging Face.
Among the Semiconductor vendors, NVIDIA made 2023 a huge investment year – it has participated in OpenAI rivals Cohere and Inflection.ai, AI21 Labs, RunwayML and PerplexityAI. Databricks and CoreWeave are major data management and cloud infrastructure players addressing heightened demand for AI workloads. Intel has taken a muted approach to AI investments - funded several of the highest-valued startups such as Hugging Face and AI21 Labs. Qualcomm, Dell Technologies, AMD, HPE, and Oracle, all have been more selective in their commitments - for example, Aleph Alpha's recent $500M round led by HPE.
Application SW vendors have also been splashing cash in the generative AI startup sphere. Salesforce has invested in unicorn Mistral AI's series A for the open-source large language model (LLM). SAP or Oracle have a close relationship with Cohere. Other players like Workday, Atlassian. and Zoom have built stakes in various LLM and generative AI specialists through their venture arms. The AI field remains mostly early stage with many of these startups rising from relative obscurity to unicorn status in a matter of months.
Scaling AI is Not Easy
It is becoming clear that getting the business to scale in AI is expensive, far more expensive than many AI entrepreneurs and the VCs that back them want to admit. As the computing costs surge startups will struggle even more to be competitive and we think large incumbent vendors like Facebook, Google, Microsoft, AWS, Oracle, Salesforce, Workday, ServiceNow, Splunk will dominate their respective turfs. Sam Altman’s recently floated goal of raising about $7T to make AI chips highlights the complexity of AI landscape. The infrastructure needed to build a foundational AI model is expensive and most of that value will be held by a handful of large technology companies. Significant new technology like AI can create seismic shifts in the economic landscape. It seems most of the early benefits may accrue to the incumbents which have substantial resources. This oligopoly will get worse, and the cost of doing AI business for AI startups - whether building a defensible AI first product or creating a friction less GTM to reach the end buyer, will become too high for them to survive on their own.
Generative AI startups can build their product in two ways - develop own version of OpenAI’s GPT-4 or Google’s Gemini like foundational AI model that requires hundreds of millions in investment or build on top of an existing trained AI model, which only needs tens of millions in investment and which most Gen AI startups do today. Using OpenAI as the foundational AI model incurs 100x in cost than what these startups can charge the customers for using their products. It is like you are connected to the electricity grid and there is constant consumption cost that is the biggest bucket in the solution cost items. If these startups decide to build/manage a foundational model, it becomes a risky balancing act as having access to GPUs on cloud infrastructure is becoming a race that bids up the infrastructure price. In both cases, the primary beneficiaries are Hyperscalers like Microsoft, Amazon and Google, and AI chip maker Nvidia. All the Gen AI startups are raising money from VCs and giving it to cloud companies and Nvidia. At this point, no Generative AI startup that must rent AI chips and cloud computing infra, has figured out how to run a low-cost business at scale, and in fact they can make money - only if no customer uses their AI product. Most of the flurry of new startups that jumped into the hyped-up generative AI market will likely fold or be folded into the incumbents over the next few years. Additionally, we see that most AI startups are not staffed with hard-core AI researchers, rather they are staffed with regular SW engineers that implement the engineering side of AI deployments to merely make them operational. The lack of core AI talent does not seem attractive for M&As by incumbents. In fact, most of the core AI researchers reside within those large incumbents and startups are not able to get this cream of talents. Most of the incumbents would rather just invest in startups with minority stakes just to keep a directional sense of AI landscape and avoid regulatory scrutiny of an acquisition. We think regulatory pressure will prevent outright takeovers of leading AI startups that have valuations over $1B, like Cohere, Character.ai and Inflection, leading to a few new standalone AI companies in the market, along with large incumbents, capturing majority of the value in the chain.
It would be naïve to assume that AI will disrupt every market in a way that will benefit startups. Founders building AI must be smart about how, and in what arenas, they choose to challenge incumbent players. There are several ways for and AI startups to experiment on what sticks in the market.
Strength of Innovation
Applied AI startups require founders to have deep vertical or domain expertise and insights that are not easily accessible to the competition. Deep vertical insights create a sustainable competitive advantage that the AI startup needs to refine the usage data and the product outcome. Founders need to analyze if the AI benefit is creating a sustaining innovation or disruptive innovation that can create new market and services. AI startups are better off by focusing on disruptive innovation where the value created by the product is 100x better than the existing ones in the market. An AI product that can eliminate large line items of customers’ OpEx or completely redefine the OpEx could be disruptive. Same goes for challenging the business model of an incumbent. If a startup can use AI to undercut existing business models and replace them with something else, that’s clearly disruptive innovation. If an AI startup can create a conflicting business model for a large incumbent, chances are that the incumbent will not take on the startup at the cost of their own business. This will result in a slow response from the incumbent. Because startups are more agile, often don’t have an existing business model to protect, are able to move faster, and often more willing to take on risks early on, AI for disruptive innovation makes great sense where startups can be aggressive and take incumbents head on. On the other hand, most of the large incumbents will play the sustaining innovation with AI due to their data lock-in advantage - Google and Facebook will use AI on top of their ad business. The tendency of the incumbents will be to bundle AI features into their existing solutions. AI startups playing in the sustaining innovation space will have challenges and may have to take a more collaborative approach with the incumbents, given the data and GTM power incumbents have. Or the startup could observe incumbent’s bundling strategy and try to unbundle with best of breed AI solution for specific use cases. There will be some push-pull between the incumbent and the AI startup in such scenarios. In some circumstances, an AI startup with a specific focus can deliver a better product than incumbents who are adding on AI features without much thought.
Unique GTM
We see that incumbents have tremendous advantage here in terms of reaching out to a potential AI customer for a new AI product. Unless startups first find a way to minimize reaction from an incumbent while trying to access customer, and the valuable data, incumbents will always be miles ahead. Early-stage AI startups need a unique capital efficient GTM (marketing & sales) to break through the AI startup clutter and solidify the product-market fit. An initial GTM playbook could be a collaborative approach with an incumbent that increases its product value prop, and then slowly creates new capabilities after becoming entrenched in the customer datasets. An outright challenge to the incumbent will have tremendous GTM friction leading to nowhere.
Speed to The Rescue
We believe AI transition will sweep across software, bio, and every other industry and speed could be a key advantage for startups before incumbents could adjust to AI. AI products are not going to be prefect right out of the door; startups need to quickly put out their product in the market for rapid adoption and then iterate with new sources of data as the product gets used. Founders need to quickly insert their AI products into new areas arising from this big transition before incumbents make up their minds on ways to incorporate AI into their existing offerings. Engineering teams need to experiment on their AI value prop 10X faster than what they were used to in traditional SaaS market space. More experiments offer insight into where AI can add substantial value that the customers will pay for, else that cost of renting the GPUs and Foundational model for low value tasks will bankrupt the startup. The field is new as people are learning and no one knows where and how AI can add substantial value.
Technical Edge
As we mentioned earlier, many AI startups are mostly staffed with traditional SW engineers. But AI startups don’t need have all PhDs in the team all the time (although having such a team increases the chances of success), but need a team that is strategic about SW Architecture, Model Selection, Ops, and Tooling while keeping the focus mainly on delivering an excellent product that drives significant benefits to end-users and can iteratively become more valuable with more usage + additional users. Additionally Human-in-Loop should be considered a part of the technical strategy. An AI product that is augmented with human expertise will drive faster adoption, frequent usage, varied end results, build trust, and achieve wider industry penetration. An ease of use and a human-centered AI product has a high chance of succeeding. We know from the past that technical edge will erode at some point, maybe after a long period if a startup has good tech DNA, but it will erode. This beautiful aspect of technology came in battery manufacturing, semiconductors, Blockchain, CPUs, Genome sequencing. And in such a final scenario, it makes sense for startups to think about multiple sources of differentiation to prolong their differentiation till the exit.
Quality of Market
We prefer AI startups that are truly disruptive in nature. Such startups do not need to worry about incumbents much if their product is unlocking something fundamentally new that was previously not there in the market. AI startups operating in a market where their fast execution speed can be advantage stand a higher chance of scaling. Incumbents inherently have data and distribution moats, and they will use it to launch AI products. But there are some markets that are usually laggards with incumbents launching products at slower pace relative to other markets. AI startups operating in laggard markets such Healthcare, Legal, Industrial, Manufacturing, Construction, Logistics have a better chance to succeed. It would be foolish for AI startups to take on incumbents in fast moving markets such as Finance, Technology or Retail where incumbents are used to bringing solutions faster. With their distribution and AI as sustaining innovation, they can easily wipe out the startups. Choosing the market wisely is key for AI startups. A multi-billion dollar massive market that can be created with a cleverly planned beach head product, followed intentionally by second + third acts of the startup that can expand TAM with additional customer segments or additional use cases.
We see many AI startups that are pretending to be something they are not in terms of value proposition – we see that many are actually pitching sustaining innovation with limited defensibility within their chosen industry – not true disruptive products. They will be out competed by incumbents with their GTM and data advantages. We also see startups trying to compete brute force in areas where incumbents hold strong advantages by trying to do what incumbents already do, but better. In these circumstances, the startups have to make sure that AI change this industry in a meaningful way that never existed better. The above are some of the strategies AI startups can deploy to win, but we are confident there will be new ways of winning by AI startups that will be clear after winners are established.
/Service Ventures Team
Comments