Deep Diving on AI Commercialization
Table of Contents I. Key Takeaways
1. Current market leaders have cost-effective, diversified monetization schemes 2. Large-scale partnerships, often backed by equity investments, provide concept validation 3. A new trend among SaaS leaders to buy and package has emerged 4. Growth equity investors stand to benefit from the barbell-shaped market 5. Horizontal AI tools favor user-based pricing (for now) 6. AI data and compute “picks and shovels” critical to broader industry commercialization 7. Current sector-oriented investments less focused on monetization timelines 8. GTM approach critical when commercializing AI-powered hardware/physical assets 9. With generative AI, value determined by metrics, not revenue 10. Open-source could expedite AI commercialization
This document has been prepared by Lazard & Co., Limited ("Lazard") solely for general information purposes and is based on publicly available information which has not been independently verified by Lazard. The information contained herein is preliminary and should not be relied upon for any purpose. No liability whatsoever is accepted, and neither Lazard nor any member of the Lazard Group (being Lazard Ltd and its direct and indirect subsidiary and associated undertakings) nor any of their respective directors, partners, officers, employees, representatives or other agents is, or will be, making any warranty, representation or undertaking (expressed or implied) concerning the accuracy or truthfulness of any of the information, ideas, forecasts, projections or of any of the views or opinions contained in this document or any other written or oral statement provided in connection herewith or for any errors, omissions or misstatements contained herein or for any reliance that any party may seek to place upon any such information. Nothing contained in this document constitutes, or should be relied upon as, (i) the giving of financial, investment or other advice by, or the issuance of research by, Lazard, or (ii) a promise or representation as to any matter whether as to the past or the future. Lazard undertakes no obligation to provide the recipient with access to any additional information or to update or correct any information contained herein. Interested parties should conduct their own investigation and analysis of the matters and companies referenced herein. Nothing contained in this document constitutes, or should be deemed to constitute, an offer or solicitation for the purchase or sale of any security. You undertake to keep this document confidential and to not distribute it to any third party, or excerpt from or reproduce this document (in whole or in part), without the prior written consent of Lazard. Lazard, which is a regulated financial adviser, only acts for those entities whom it has identified as its client in a signed engagement letter and no-one else and will not be responsible to anyone other than such client for providing the protections afforded to clients of Lazard nor for providing advice. Recipients are recommended to seek their own financial and other advice and should entirely rely solely on their own judgment, review and analysis of this document. Lazard or other members of the Lazard Group (i) may have acted in the past, or act currently or in the future as adviser to some of the companies referenced herein, (ii) may receive fees in connection with any such advisory engagements, (iii) may at any time be in contact with such companies in order to solicit them to enter into advisory engagements, and/or (iv) may from time to time have made, and may in the future make, investments in such companies. By accepting this document, recipients agree to be bound by the terms and conditions set out above.
Section I: Key Takeaways
We evaluated pricing and go-to-market (GTM) models for 150 of the most promising VC-backed AI companies, ranging from pre-revenue to unicorn-stage innovators. Here are our top 10 learnings:
1. Current market leaders have cost-effective, diversified monetization schemes – based on the subset of late-stage growth ($40-100M estimated revenue) and scaled ($100M+ estimated revenue) AI companies we examined, today’s market leaders have maximized their commercialization capabilities through layered pricing mechanisms and GTM strategies. These companies averaged almost 3x the number of combined pricing models and third-party sales channels leveraged compared to their early- and mid-stage (<$40M estimated revenue) peers. We also found that hybrid GTM models encompassing both user- and usage-based/pay-as-you-go pricing, often coupled with feature add-ons or additional products offered at flat or tiered rates, were most correlated with higher revenue scale and market value (most recent post-money valuation). In traditional SaaS, usage-based models have been associated with some of the highest-value market leaders, particularly in the infrastructure category (i.e. Snowflake, Datadog, Zscaler).
Because gross margins of high-performing AI software companies (~50-60%) often trail traditional SaaS benchmarks (~75-85%) – primarily due to higher input costs and the need for customer-specific services – hybrid subscription and consumption-based models can improve overall margin profiles by encouraging upsells while also providing flexibility in onboarding new customers who are still in an AI discovery/experimentation phase. AI software can generate substantial upside value through data-consumption monetization models – especially when layered onto flat-fee access subscriptions – as data volume and quality are direct determinants of AI system success and ROI. We think AI companies that adopt hybrid pricing strategies as part of their initial GTM strategies will have a competitive advantage in capturing maximum value from early, flagship customers.
Figure 1: Benefits of hybrid monetization models to maximize customer value – SaaS sample
Usage-based companies have better NDR...
...but also lower gross margins
NDR = Net dollar retention
Sources: Lazard VGB Insights, a16z, Bain & Company, OpenView Partners
2. Large-scale partnerships, often backed by equity investments, provide concept validation – the natural extensibility of AI’s core value propositions – automating manual processes, personalizing customer experiences, and making predictive interpretations of the data flowing through a hardware/software platform – enable AI providers to reach a much broader audience by working with channel partners and third-party marketplaces. This trend held true in our analysis, as the monetized partnerships category was the greatest differentiator between the “growth-scaled” and “early-mid” staged companies we examined, as shown in Figure 2. Additionally, we found that growth-stage companies more frequently offered add-on features sold independently, activating expansions into their sales motions. Early- and mid-stage counterparts, in contrast, relied heavily on freemium, tiered monthly rates – mostly based on team size and use case – with comparably simpler product/feature packaging.
Figure 2: Greatest monetization category disparities – early-mid vs. growth-scaled AI companies
Source: Lazard VGB Insights
A success story leveraging this model in its early growth phase was Databricks, which used partnerships and integrations, along with new feature development, to scale from an open-source project to an at-scale industry leader with over $100M in ARR in just three years. The company partnered early with Microsoft—who also became a strategic investor—to collaboratively develop the Azure Databricks offering and enable 700M+ Azure customers to consume their products without friction through the Azure Marketplace. Databricks also deftly took a cloud-agnostic approach in its GTM strategy by enabling customers to pay compute and data storage costs directly to the cloud providers rather than collecting this as direct revenue, which established friendly dynamics with the hyperscalers. Additionally, in its quest to further differentiate from Snowflake as it scaled to over $1B in ARR, Databricks continuously layered new features into its enterprise platform to become the pioneering “data lakehouse”—converging many of the capabilities of a data lake and warehouse into one.
We’ve also seen this approach validated at the infrastructure layer through equity commitments and strategic partnerships executed by leading cloud/SaaS vendors. These tech giants are operating under the thesis that they can create an “economy of scale effect” to commercialize AI by leveraging their existing GTM engines, deep balance sheets, and computing resources. Rather than build in-house, current market trends suggest partnerships and integrations are widely viewed by legacy vendors as the best entry point for upscaling AI research and development, creating purpose-built models, and incorporating AI capabilities into existing commercial-ready consumer and enterprise products.
Figure 3: Select cloud/SaaS leaders’ AI/ML infrastructure investment and partnership activity
Sources: VentureBeat, Financial Times, TechCrunch, Company Websites, Press Releases, Tomasz Tunguz, Contrary Capital
3. A new trend among SaaS leaders to buy and package has emerged – despite the proliferation of partnerships dominating the market leaders’ growth strategies thus far, we have recently seen enterprise software companies pay upfront – often at a premium or flat value to a recent highly-priced equity raise – to acquire and integrate AI/ML solutions into their existing platforms with the stated intent of quickly commercializing a bundled offering (see Figure 4). The dynamic nature of today’s AI foundational models and data infrastructure – which serve use cases across nearly every industry and attract both tech and non-tech users – presents a challenge for B2B software companies seeking defensible AI/ML strategies. This fight to establish a competitive moat has accelerated M&A timelines and set a foundation for valuation multiples that is highly inconsistent with current market precedents. Buyers increasingly seek to own powerful model/infrastructure assets and industry-oriented applications before they have gained proven commercial traction, believing their existing breadth of GTM capabilities will create an “economies of scale” effect and enable them to be first-to-market with pre-packaged solutions specific to their customer base.
Sources: Richard Waters (FT), Lazard VGB Insights
Figure 4: Select AI/ML infrastructure M&A activity – 1H 2023
Sources: Transaction Press Releases, Pitchbook Data, Inc.
4. Growth equity investors stand to benefit from the barbell-shaped market – disproportionately significant volumes of early-stage AI companies are flooding the market, largely enabled by the model and infrastructure providers who have borne the high costs of model-building and data-tuning to facilitate rapid, low-cost shipment of new AI/ML applications. This nascent, yet fast-growing product development activity – coupled with a highly concentrated funding environment – has created a barbell dynamic in the AI/ML market (see Figure 5). We believe this is likely to create a wave of investment opportunities at the Series B and C stages over the next 6 – 12+ months as more sector-specific AI enterprise tools will attract more generalist B2B SaaS growth investors into the AI funding race. The most active early-stage AI investors to date – including brand names like Sequoia, Index Ventures, a16z, Tiger Global, and Khosla – will provide a validated pipeline for top-tier growth investors to mine, while the bulk of mid-quartile growth investors will need to follow Softbank’s lead in crafting separate AI mandates enabling them to invest earlier than in other verticals.
Figure 5: Barbell dynamics of the AI market – historical deal counts and capital invested ($M)
Deal counts reflect flurry of new company formation
Capital highly concentrated among select winners,
Series B/C financings yet to take off
Source: Pitchbook Data, Inc.
>50% of total AI funding through 1H ’23 sourced from
7 late-stage infrastructure deals
Funding data displayed in Figure 6 suggests that the tier 1 VCs investing at the early stages are also looking beyond current market conditions when underwriting new AI investments, adopting a long-term view that backing potential category leaders early amid a new platform shift justifies premium pricing, even at the risk of incurring higher-than-average loss ratios. In contrast, median late-stage deal sizes have followed a lumpy trajectory as investors have recently focused on the 20 – 40 top-funded AI/ML infrastructure providers. Median deal sizes contracted significantly in Q1 ’23 to sub-pandemic levels before doubling back to pre-COVID norm levels in Q2.
Figure 6: Early-stage AI funding trends suggest investors are taking a long-term view
Investment into early-stage AI is countering broader market trends…
Sources: Carta Insights, Pitchbook Data, Inc.
…and these companies are attracting higher pre-money valuations
Late-stage AI median deal size fluctuations reinforce a nascent, concentrated market for growth investors
Sources: Carta Insights, Pitchbook Data, Inc.
5. Horizontal AI tools favor user-based pricing (for now) – our analysis found that industry-agnostic AI software – encompassing general enterprise workflow tools and generative capabilities to deliver and enhance form content (text, image, and video) – skewed heavily towards seat-based pricing methodologies by a factor of 3x over others included in our sample. Nearly 80% of all GTM strategies centered around user-based pricing that we studied came from this horizontal subset. There are several potential explanations for this trend: 1) usage-based models are more challenging to implement as companies need to reward high-volume consumers while also finding ways to drive higher engagement among low-average users; 2) generative tools have myriad use cases for content development with unclear ROI distribution to inform which usage metrics to incentivize through pricing schemes; and 3) general back-office/ERP solutions often have standardized user profiles with limited upsell potential on a per-user basis. This is consistent with Bain & Company’s recent findings on horizontal applications being a relatively poor fit for usage-based models relative to infrastructure platforms that leverage data as the core asset.
Sources: VentureBeat, Lazard VGB Insights
Figure 7: Share of customers using consumption models vs. those who want to use them (2022)
Source: Bain & Company Technology Report 2022
6. AI data and compute “picks and shovels” critical to broader industry commercialization – enterprises are rushing to identify differentiated ways to incorporate LLMs and vector databases into their tech stacks, and as recently noted by NVIDIA CEO Jensen Huang, are increasingly focused on cloud-first AI strategies that enable fast development and scalable deployment. The quickest scalers that we studied in the AI/ML infrastructure category all had a common trait of being the early facilitators—"the picks and shovels”—enabling enterprises to leverage models with their own proprietary, unstructured data and access necessary compute power to build scalable applications. While much of the industry’s focus has been centered around the model providers themselves, the data and resource optimizers are the ones most impacting commercialization across the broader market.
Scale AI is one example of a company that has maximized its market value (>$7B) by being the go-to platform that sits between the raw data and the AI models themselves, acting as an enabler for companies seeking to leverage smarter AI capabilities but without the technical resources to implement them. The Scale AI platform automates the manually-intensive process of annotating and labeling enterprise data before it can be fed into AI models, and through its own rigorous back-end ML model training, is able to do so in smarter ways than if humans controlled the process. Having grown from an image and video-tagging business in its early days, the company has expanded its GTM strategy over time by focusing on volume-based pricing that scales with the data labelled for customers (sticky expansion opportunity) and adding offerings such as data debugging tools and synthetic data generation to fill in gaps from customers’ existing datasets. Being the foundational access vector and democratizer to AI models has enabled the company to amass significant market share and fend off competition from earlier-stage players such as Labelbox and DataLoop.
On the compute side, Coreweave has differentiated itself by being the first at-scale access provider to NVIDIA GPUs (highest quality for AI models), and claims to do so at 80% less cost than existing cloud providers. This has led to large-scale, monetized strategic partnerships, including a recent deal with Microsoft that is reported to be worth billions of dollars over multiple years. The company has grown its 1,000+ customer base across four verticals: generative and open-source AI/ML, batch processing, pixel streaming and visual effects, and rendering. Even while competing head-to-head with the major cloud providers – AWS, Google Cloud and Azure – Coreweave has successfully marketed itself as the leading hardware provider specifically for inference of AI models. Ultimately, Coreweave’s early GTM strategy of offering access to best-in-class GPUs at customer friendly, usage-based rates has been the differentiating factor enabling its scale. Whether the company can maintain that pricing advantage, or diversify its generative AI-led customer base, will ultimately determine its growth potential in a crowded market.
Sources: Wall Street Journal, Contrary Capital, Company Websites, Lazard VGB Insights, CNBC, TechCrunch
Figure 8: Select AI data and infrastructure providers
Sources: Crunchbase, Pitchbook Data, Inc.
7. Current sector-oriented investments less focused on monetization timelines – our sample found that vertical-focused AI deals to date have looked more like DeepTech investments rather than traditional vertical B2B software plays, judging from the risk profiles, long lead-times, and uncertainty around customer adoption inherent to these companies. We believe this dynamic will evolve as the application layer continues to be built out and as industry-specific models enable more widespread integration of AI functionality onto enterprises datasets. Healthcare, mobility, InfraTech (industrial + logistics/supply chain), and clean energy technologies were the predominant vertical-focused solutions covered in our analysis. The common traits of these businesses include:
- High barriers for market entry with required trials/proof-of-concepts
- Proprietary, self-generating data sets – often with a hardware component
- High CAPEX requirements for product development and to reach operational scale
- Industry-specific regulatory and customer adoption hurdles
- Longer relative sales cycles, though often bringing long-term, high-upside contracts
Figure 9: Industry distribution of vertical-focused AI companies in our sample
Source: Lazard VGB Insights
Despite this, we think a new wave of verticalized AI applications that look more “SaaS-like” – automating more traditional B2B workflows specific to an industry – is likely to flood the market in the next 12 months. Rather than being developed to solve complex technical problems or enable novel product creation (i.e. new medical therapeutics, innovative infrastructure projects), entrepreneurs are likely to focus on delivering solutions trained on highly-specific datasets that provide instant ROI to industries with content-based workflows at the center of their core operations – this is why the legal industry has been a proving ground for verticalized AI to date. However, in the near-term, more economic value will likely be created by incumbent SaaS leaders integrating AI functionality into their existing offerings. Index Ventures’ Paris Heymann’s recent vertical AI analysis outlined several factors that could determine which of these two strategies commercializes faster:
- New AI-native vertical applications have a natural wedge against traditional SaaS in automating manual processes and delivering quick time-to-value, particularly in knowledge work industries
- Existing vertical SaaS companies with extensive customer data sets and robust distribution channels can more easily layer AI functionality into existing products, and risk falling behind the competition if this is not made a top strategic priority
- Ultimately, pricing new native vertical applications will be a challenge until customer value can be accurately forecasted, whereas legacy vertical SaaS companies can quickly package new AI features into existing product packages or offer them as tiered add-ons
Sources: Lazard VGB Insights, Index Ventures
8. GTM approach critical when commercializing AI-powered hardware/physical assets – the AI-powered hardware and physical asset providers we analyzed were among the most creative in designing GTM strategies. Despite facing more complex sales cycles landing large customers such as government agencies or major health systems with AI-powered hardware leases or sales, these companies benefit from being able to collect and analyze new types of data through the deployment of physical assets. Coupled with AI-powered software that integrates existing customer data, these companies have an advantage in bringing personalized AI solutions to market. While often bearing significant upfront CAPEX for product development, asset providers often sign multi-year, custom master services agreements (MSAs) with individual customers that offer higher average values than are seen with typical annual software licensing subscriptions. However, given the sensitivities around automated physical asset deployment in industries like defense and healthcare – where lives are potentially at stake – there are often extended lead times to commercialization.
That said, creative GTM approaches have enabled some industry leaders to circumvent these hurdles, including autonomous aircraft and weapons manufacturer Anduril. The company scaled to $10M in revenue in its first 22 months, a milestone which took SpaceX five years to reach, and one Palantir struggled to surpass while selling to the same defense industry customer base. Anduril learned from its industry peers’ experiences to accelerate the company’s sales motions into the federal government, averaging a much quicker 9 – 12 months to build and ship minimal viable products, and centering its sales around the LatticeOS software platform (“the brain”) that powers and integrates the various sentry towers and autonomous aircraft systems it sells. The business model of building “software-native defense systems”—hardware that is operating almost entirely on intelligent software—and selling its software as the initial “land” to later “expand” with its physical assets, has differentiated the company from other field asset providers and enabled the company to scale above $150M in revenue as of 2022.
Figure 10: Select AI hardware/software provider GTM models
Sources: Company Websites, The Generalist, Lazard VGB Insights
9. With generative AI, value determined by metrics, not revenue – lofty generative AI valuations relative to the broader market is not a new trend, however the widening delta since pre-2021 has illustrated investors’ flight to identify the most promising early-movers (see Figure 11). Through 1H ‘23, AI startups comprised over half of all VC-backed unicorns newly minted this year.
Figure 11: Median early-stage pre-money valuations ($M) – generative AI vs. all startups
Source: Pitchbook Data, Inc. “In the world of startup valuations, there's generative AI—and everything else”
Our sample of generative startups included >90% of companies at the early-mid stages in terms of revenue traction. The nascency of the market, coupled with the consumer-like nature of their business models (upfront focus on user acquisition), has made revenue an almost obsolete metric in terms of assessing a company’s current and future value. So how are investors evaluating “traction” in the generative AI space? Our analysis illustrated the following insights:
Figure 12: Sample generative AI companies’ commercial traction metrics
Sources: Pitchbook Data, Inc., Funding Press Releases, TechCrunch, Company Websites
10. Open-source could expedite AI commercialization – based on the data we could gather, at least 20% of the AI companies we examined had inputs and/or development roots from open-source projects; what’s notable about these startups is that the majority have quickly adopted enterprise monetization playbooks by introducing “freemium” tiering that incorporates flat-fee or user/usage-based subscriptions to sell premium features. As noted by CRV’s recent analysis on open-source in AI, historical precedents show the majority of thriving open-source companies take 2-3+ years before introducing monetization strategies (i.e. Confluent, MongoDB, Elastic). With AI, free open-source LLMs like Stanford University’s Alpaca can leverage the outputs of closed-source models (ChatGPT) to offer enterprises a quicker on-ramp to AI models and integration into existing products.
A recently leaked internal Google document validated that the company’s senior engineers are concerned about the rapid progress made from their open-source competition. The letter noted that open-source models are faster to develop, more customizable, more private, and as – if not more – capable. Open-source models are also free, unrestricted, and equally friendly to non-tech users, creating the potential for them to play a central role in democratizing AI’s commercialization within the enterprise. Regulatory challenges – coupled with bottlenecks created by resource scarcity (compute, engineers) and cloud provider relationships among closed-source competitors – will ultimately determine whether open-source models and infrastructure can maintain a steady pace of growth and adoption.
Sources: The Guardian, CRV, Lazard VGB Insights
Figure 13: Historical % open-source vs. closed-source by technology type
Source: Kelvin Mu (Translink Capital), Lazard VGB Insights
Section II: Methodology / Sample Details
Our sample of 150 AI-centered companies was generated by names primarily included in the Forbes AI 50, CB Insights AI 2023, and NFX Hot AI 75 lists—all of which were released in Q2 '23 and were compiled by credible AI investors and technical experts. Before jumping into the actual sample data, we used the following criteria and definitions to segment the market:
AI/ML Infrastructure – Includes foundational models (i.e. Anthropic, Cohere) and other “picks and shovels” like databases and model-tuners/enhancers
Application: Horizontal – AI application software (including generative) offering an industry-agnostic functional capability
Application: Vertical – AI application software (including generative) catering to a specific industry use case
Enterprise Infrastructure – AI-powered enterprise solutions for security, data, compute, and other infrastructure needs
Hardware/Application: Horizontal – AI-powered hardware/physical assets, often coupled with software, delivering value to customers across industries
Hardware/Application: Vertical – AI-powered hardware/physical assets, often coupled with software, delivering value to customers in specific industries
Sample Breakdown by Category, Pricing Models, and Estimated Revenue Stage
While heavily skewed towards companies in early stages of development and commercialization, the sample was designed to reflect the nascency of the broader AI market, which includes ~125 unicorns in total. The following Figures 14 - 16 provide further details on the distribution of companies sampled by AI category, pricing model, and estimated revenue stage:
Figure 14: Breakdown of sample by business model and estimated revenue stage
Figure 15: Sample breakdown by AI category
Figure 16: Monetization strategies used by sample participants by AI category
Sources: Pitchbook Data, Inc., Funding Press Releases, TechCrunch, Company Websites