2024 Year-In-Review + 2025 Predictions

In this last post of the year, I look back at five of the biggest B2B AI & software debates in 2024, and preview what 2025 may have in store for us.

Building on momentum from 2023, AI continued to dominate the venture and tech landscape in 2024. This year, as the industry rapidly matured beyond the concept of LLMs which initially captured the world’s attention just a few years ago, AI discussions moved towards second order questions around applicability and implementation. Given the fast pace of progress, it’s been challenging to summarize all the exciting developments within the past 12 months. So I’ve leveraged my most-read posts of the year to reflect on 5 of the biggest B2B AI and software debates in 2024 and offer corresponding predictions for 2025:

1. AI companies are seeing real revenue traction, but moving from POC to prod has been slower than expected

2. AI agents and the “Death of SaaS”

3. LLMs potentially hitting a scaling wall and the quest for new model development paradigms including “reasoning”

4. A new era of AI-generated content with 2024 being AI video’s banner year

5. Tip of the “Iceberg” for data interoperability

---

I. AI companies are seeing real revenue traction, but moving from POC to prod has been slower than expected

2024 was a year where AI companies saw real revenue traction (chart below), cementing the fact that AI is no longer an aspirational concept, but is now a very tangible one where consumers and enterprises are willing to spend dollars on. The revenue growth of some AI companies has been truly astounding, leading investors to assign premium multiples to these companies.

Source: Sapphire Ventures, “Top 10 AI Trends & Predictions for 2025” (12/17/24)

However, while many AI companies demonstrated strong revenue traction this year, it’s important to note that portions of this revenue remain as part of non-recurring experimental budgets, since the graduation rate of proof-of-concepts to production deployments seems to be lower than previously expected at the enterprise level:

Source: Battery Ventures, “State of the OpenCloud Report” (November 2024)

Frictions to widespread adoption within enterprises include factors such as a lack of use case applicability, unproven ROI, and data privacy concerns (chart below). These issues slowed AI’s transition from the realm of novelty to true necessity within organizations.

Source: Morgan Stanley, “3Q24 CIO Survey” (10/10/24)

Prediction #1

Many stakeholders are currently working on addressing various adoption hurdles. While I don’t expect all the friction points to be fully resolved in short order, I’m optimistic that more progress will be made in 2025 such that the next year could herald an inflection point where a majority of enterprises will have at least one AI application deployed in production:

Source: Morgan Stanley, “3Q24 CIO Survey” (10/10/24)

---

II. AI agents and the “Death of SaaS”

AI agents became all the rage in 2024, especially as enterprise leaders focused on productivity use cases for AI deployments:

Source: Morgan Stanley, “3Q24 CIO Survey” (10/10/24)

Unlike a simple “assistant”, AI “agents” go one step further to demonstrate capabilities and advanced reasoning that enable them to perform tasks on behalf of users. This disruptive potential of agentic AI was a primary factor in the emergence of a “Death of SaaS” narrative throughout 2024. Satya Nadella pushed this debate again this month as he posited that traditional SaaS apps would be reduced to glorified databases as AI agents take over the business logic layer.

Prediction #2

I believe that AI agents will usher in a reimagining of how human users interact with software, pushing the boundaries on our conventional view of a “GUI”. Additionally, AI agents will not only disrupt how human users interact with software, but will also impact how machines interact with machines. From computer use to LLMs.txt, I anticipate 2025 being a year where more novel UI/UX mechanisms are developed for AI-native workflows as software interfaces adapt to an agentic AI world.

---

III. LLMs potentially hitting a scaling wall and the quest for new model development paradigms including “reasoning”

Debates around LLM performance reaching a plateau dominated AI discussion circles in 2024, with AI leaders divided on the future of scaling laws. The scaling debate has mostly been centered around how using more data and compute power in the LLM pretraining stage may not necessarily continue to result in the level of commensurate performance gains seen in the past. Ilya Sutskever further fanned the flames of this debate at NeurIPS this month by boldly declaring that “pre-training as we know it will unquestionably end”.

In response to LLMs potentially hitting a scaling ceiling, researchers began actively pursuing new model development paradigms. In 2024, we saw OpenAI lead the way by introducing state-of-the-art “reasoning” models, such as o1 and o3, that emphasized time-test compute with inference-time scaling:

Source: Tweet from @DrJimFan

Early “reasoning” models are impressive and are a step-up from standard LLMs for many applications, but they are still bound by limitations of transformer-based architectures. Thus, the search for new paradigms has also pushed researchers to innovate on new model architectures beyond the transformer. In fact, one of my predictions for 2024 was that it would be a year where we see more non-transformer based approaches emerge as game-changing contenders. This has come to fruition with examples of SSM pioneer Cartesia and “world model” advocates such as World Labs making significant waves in 2024.

Prediction #3

As 2024 has shown, the model layer remains as one of the most vibrant and innovative parts of the AI infrastructure stack. On the technical side — I believe we’re at the start of a very exciting era for “reasoning” models and I’ve written about the potential near term impact that these models could have on the applied AI landscape.

On the business and product side — 2025 should be another dynamic year for the foundation model layer wars. Earlier in 2024, my colleagues and I discussed different frameworks for understanding value capture within the model layer. As discussed in Section I, many businesses within the model layer have experienced hypergrowth over the past year. There are now considerable talent and capital barriers to entry at the model layer. In 2025, I anticipate some consolidation to occur within the model layer as a handful of emerging leaders — OpenAI, Anthropic*, xAI, Meta, and Google — continue to pull away from the pack and establish themselves as the new AI Fab 5. Each will likely cement their position with their unique approach and distribution advantages — for instance Meta with open-source and Anthropic* with AI safety. While initially slow to react, Google has emerged as a surprisingly strong contender this year, launching a slew of new AI releases in the last month alone.

---

IV. A new era of AI-generated content with 2024 being AI video’s banner year

Source: Nathan Benaich and Ian Hogarth, “State of AI Report 2022” (10/11/22)

Last year, I predicted that multimodality will become table-stakes. At the time, video modalities made up only a very small percentage of AI research (chart above). Momentum shifted in 2024 which marked a banner year for Video AI; notable developments this year included:

• The release of Google DeepMind’s Veo 2, OpenAI’s Sora, Stability’s Stable Video Diffusion, and Meta’s Movie Gen.

• Video AI startups such as Haiper, Luma, Moonvalley*, Pika, and Runway, all had new flagship model releases and/or cutting-edge feature launches this year; I’ve written previously about my excitement for the Video AI market and how startups can lead innovation on this front through differentiated approaches.

• Video AI innovation became global in scale, with Chinese players Kling AI and Minimax launching impressive models this year.

Prediction #4

2023 was AI Image’s year, 2024 was AI Video’s year, and I anticipate 2025 to be when Voice AI takes a big leap forward. As my colleagues pointed out in Bessemer’s Voice AI roadmap, there is no better time for AI founders to be building in the voice modality given recent tailwinds such as a growing array of high-quality infrastructure to support the conversational voice stack, as well as groundbreaking progress with the rise of Speech-To-Speech models.

Source: Bessemer Venture Partners, “Roadmap: Voice AI” (11/11/24)

---

V. Tip of the “Iceberg” for data interoperability

Data and AI are inextricably linked. Within the data infrastructure world, buzz for data Lakehouse architecture has been brewing for a while now. Core to the rise of Lakehouse architecture has been an on-going debate around which modern open table format — Delta Lake, Iceberg, or Hudi — will emerge as the dominant standard.

2024 marked a defining year for Iceberg in the modern table format wars. On the community side, adoption across all open table formats ramped quickly, but Iceberg experienced the fastest growth (chart below). Influential startup leaders have also rallied strongly around Iceberg this year. Additionally, many major incumbents including Snowflake, Amazon, and Microsoft, announced support and/or dedicated features for Iceberg this year. And perhaps in the biggest testament of all, Databricks (that had previously created Delta Lake) acquired Tabular, the original creators of Iceberg, in a landmark deal over the summer.

Source: Morgan Stanley, “What's New in the New Stack?” (6/11/24)

Prediction #5

Excitement for Lakehouses has grown over the years primarily because this type of architecture addresses the limitations of data warehouses and data lakes, with the potential to unlock greater interoperability since different query engines can be connected to a unified open table format. This could usher in sweeping changes to the ecosystem such as unbundling and reduction of vendor lock-in. Many industry players have acknowledged and addressed this theme in 2024, and I believe we are just entering the age of interoperability for data workloads.

I also anticipate that convergence toward Lakehouse architecture with Iceberg at the core will cause an evolution in the data infrastructure stack, creating new categories but also disrupting established ones. Stay tuned for a deep dive post on innovations within the Lakehouse stack!

---

Moving beyond LLMs

The AI industry is maturing very quickly beyond LLMs. We are now pushing toward new frontiers, both in terms of technology as well as on the business front, as we seek answers for second order questions around monetization models, scaling, applied AI, sophisticated multimodal use cases, and data interoperability. More questions (and hopefully answers) will be in store for us in 2025 as AI innovation continues to accelerate at an unprecedented space. Wishing everyone a peaceful and joy-filled new year!

*denotes a Bessemer portfolio company

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of her employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.