State of the Cloud 2024

In last year’s State of the Cloud report, six months after the launch of ChatGPT, we declared that language model powered AI was here to stay and would impact nearly every cloud roadmap. In hindsight, this was the understatement of the year. Even our optimistic take underestimated the accelerated technological change that followed in the past 12 months (a.k.a. 12 “AI years”).

In 2023, our enthusiasm in AI was underscored by our committed $1 billion in funding. A year later, we have total conviction in the promise of AI — we’ve backed multiple new investments in AI startups and the broad adoption of LLM technology in nearly every portfolio company roadmap. It’s rare to find a cloud company these days that isn’t, at some, level an AI company.

We’re not alone, of course, in our enthusiasm. New technology waves often whet VC appetites, but the speed of VC reaction to this wave is wild compared to historical precedents. While The BVP Nasdaq Emerging Cloud Index (EMCLOUD) remains down from ZIRP highs and trades at historical norms, the private sector has rebounded and arguably bubbled up again, largely on the back of AI Cloud.

And AI is all we can think about. Yes, there are still meaningful opportunities for the legacy cloud with embedded payments and payroll continuing to open seams in new verticals. These verticals include underserved sectors like supply chain and logistics, trucking, govtech, and climate, which are seeing signs of life.

And yes, there is drama watching cloud companies that were overvalued and overfunded in the Covid bubble crash come down to earth or pivot to austerity.

But as we surveyed our 62 global investors we weren’t surprised that 124 eyes were trained on what we collectively agree is the largest shift in technology in our lifetimes.

This year’s State of the Cloud report serves both as a eulogy for the legacy cloud and a celebration of the current AI Cloud moment. In the words ahead, we focus entirely on the top five trends deemed most relevant to the AI Cloud by our global team of investors and the implications for the thrilling years we predict are headed our way.

For State of the Cloud 2024, we delve into five of the most powerful trends shaping the future of the AI Cloud economy and our predictions for what to expect by 2030.

The five trends defining AI Cloud in 2024

Trend 1: AI foundation models set the stage for Big Tech’s new battle-of-the-century

When we reflect on the platform shifts of yesteryear — from Browsers and Search to Mobile and Cloud — every technological change catalyzed competition to control the foundational layer. The age of AI is no different. Foundation models are the new “oil” that will fuel downstream AI applications and tooling.

In 2023, foundational model companies captured the lion's share of venture funding in AI, accounting for over 60% of total AI dollars. Players such as OpenAI, Anthropic, Mistral, Cohere, among others raised $23 billion at an aggregate market cap of $124 billion, underscoring their critical role in the AI ecosystem globally. Notably, this influx of capital was primarily driven by corporate VCs, who represented 90% of private GenAI fundraising in 2023 (up from 40% in 2022 according to Morgan Stanley). Big Tech companies such as Microsoft, Google, Amazon, NVIDIA, and Oracle, now have significant stakes in foundational model companies, as these investments are strategically aligned to enhance the AI capabilities of these tech giants, driving consumption of their core cloud and compute services. This is in addition to Big Tech companies working on their own in-house foundation model initiatives, such as Google’s Gemini and Meta AI’s Llama.

With so much funding flowing into this fundamental layer, competition is intensifying at an unprecedented pace, driving an incredible amount of innovation in the ecosystem. Here are some key trends we’ve observed in 2023:

• Base models improving rapidly: General purpose LLMs are getting better and better every second, both in terms of base performance capabilities (such as accuracy and latency) but also at the frontier including multi-modal functionality. The launch of GPT-4o left all our jaws dropping — the new release demonstrated the capability to view and understand video and audio from an uploaded file, as well as generate short videos. The dizzying pace of model improvement raised obvious questions around investment strategy in models that seem to have a half life measured in months.
• Battle between open and closed source intensifies: As we touched upon in last year’s State of the Cloud, the open source vs. closed source debate remains a hot topic in 2024 as open-source leaders closely track close-source model performance, especially with the recent launch of Llama 3. New questions have been raised around regulatory impact, whether closed-source players should open-source their older models as part of a new commercialization strategy, or if this might be the first time in history where an open-source leader might become the winner of this market.
• Small model movement gets big: Additionally, 2023 also saw the rise of the small model movement, with HuggingFace CEO and Co-founder Clem Delangue declaring 2024 will be the year of SLMs. Compared to larger counterparts, examples like Mistral 8x22b which launched this year have shown that bigger isn’t always better in terms of performance, and that small models can have significant cost and latency benefits.
• Emergence of novel architectures and special purpose foundation models: In 2023, we also saw excitement around the emergence of novel model architectures beyond the transformer, such as state-space models and geometric deep learning, pushing the frontier on foundation models that can be less computationally intensive, able to handle longer context, or exhibit structured reasoning. We also saw an explosion of teams training specific-purpose models for code generation, biology, video, image, speech, robotics, music, physics, brainwaves, etc., adding another vector of diversity into the model layer. We discuss this trend in our recent AI Infrastructure roadmap.

With so much happening at the foundational layer, it often seems like the ground is shifting beneath our feet every second. Despite the copious amount of funding that has been invested here, there isn’t necessarily a clear consensus on the winner right now.

Prediction: The battle of AI models will remain white-hot for the foreseeable future since this is a critical “land grab” that determines which Big Tech companies reign supreme within the cloud and compute markets in the coming years.

There are three possible realities we expect to see in the foreseeable future on who will capture the most value in this model layer fight:

Reality 1: The model layer becomes commoditized.

Will hundreds of millions of dollars of capital be squandered as VCs and Big Tech back the derby of AI leaders? The most well-capitalized models does not mean they’ll become the winner, as open-source models continue to closely challenge the leading market players. But a future where AI models are commoditized doesn’t necessarily imply the value of models will diminish. AI models as commodities will be akin to compute or oil as commodities — they’ll one day become the assets essential for global business operations. In this reality, the ultimate value in the AI ecosystem will be captured by compute and cloud service providers, marketplaces, and applications — not by the models themselves. However, in a world where AI models are commoditized, as we’ve seen in the oil market, this could still give rise to one or two extremely valuable companies selling these “commodities.”

Reality 2: AI Model Giants split the pie.

Similar to the Cloud Wars, a handful of notable new model companies, heavily supported by Big Tech strategics or corporate VCs, will own the foundational model ecosystem and become giants. Each of these winners will find a differentiated wedge to pair with technological differentiation, whether that’s via distribution, price/cost efficiencies, regulatory impact, etc. There could still be a long-tail of different players (especially open-source) but value will accrue to the top handful of model players. It’s not only the superior technology that will determine tomorrow’s AI Giants, but also their established distribution. 

Reality 3: AI models become as diverse and popular as the potato chip market.

Just like there are endless flavors of potato chips, the future of the AI model economy could very well look something similar to the snack aisle at your local grocery store. Many model companies can thrive as there are enough differentiated use cases (e.g., modalities, performance, latency, cost, security, etc.) for different model companies to survive. Additionally, geography and regulation could play a role here if geopolitical considerations enter the realm of AI models, with a mix of regulations and sovereign concerns supporting the proliferation of diversity in this layer.

Prediction: While we’re far from consensus, a slight majority of our partnership predicts Closed-Source models will drive the bulk of LLM compute cycles, and AI Model Giants will eventually split the economic pie (Reality #2).

We expect to see Cloud Giants leverage their access to compute, chips, and capital to influence the battle in their favor. And the frontrunners are already in the race — Microsoft/OpenAI, AWS/Anthropic, Google/Gemini, with Meta/Llama as the Linux-equivalent OSS alternative, including Mistral as a European lead.

Deep Dive: A behind-the-scenes conversation on the AI model layer 

Trend 2: AI turning all of us into 10X developers

The modern engineer has always been part builder and part student — completing a day job while constantly working to stay up-to-date on new languages, frameworks, infrastructure, etc. The AI quake has added a Ph.D. requirement to the job, as developers face a completely new set of toolchains and best practices for leveraging constantly evolving LLMs, including a new infrastructure suite for data management, curation, prompting, pre-training, and fine-tuning. Each year in the AI era requires coming up to speed on a decade’s worth of new developer knowledge.

But AI may also offer the solution to this new level of complexity. 2023 saw widespread adoption of code copilots and the first few months of 2024 saw early breakthroughs in agentic tooling that suggest that end-to-end automation of simple code tasks and perhaps much more may be arriving sooner than we might have expected.

Prediction: The role of the developer will be radically transformed, perhaps more than any other profession, by AI. By the end of the decade, significant developer capability will be available to every human with a computer. The resulting rate of software development will melt keyboards and dramatically reduce the age of the average technology startup founder.

Three main areas are driving the lightning speed evolution of the AI developer economy:

1. The code copilot industry has been a hotbed of innovation and competition with $3.9 billion VC dollars invested in 2023 for GenAI technology and tools. Github’s incumbent Copilot product, powered by OpenAI’s GPT-4 and Codex models, is well-penetrated with north of 14 million installs. A long-tail of well-funded and scaling startup competitors, such as Tabnine, Magic.dev, Augment, Poolside, Cursor AI, OpenDevin, Cognition’s Devin, and Supermaven, are building and iterating with developers in the loop. Some, like Magic.dev, Poolside, Augment, and Supermaven, are pre-training their own large AI models with an emphasis on model properties such as context, latency, etc. Others, like Cursor, are model agnostic and are focused on the developer experience, interface, and workflows. This landscape is a good example of the capital intensity of model-layer AI companies; Magic.dev, Augment, Poolside, and Devin have each raised $150M+ in the last couple of years.

2. The “graduation motion” of copilots embedding agentic search and generation functionality will drive outsized value in the coming years. Devin, SWE-agent and OpenDevin have demonstrated the potential of end-to-end agentic tools that interact with developer environments (i.e., file editor, bash shell) and the internet to complete coding tasks. Underpinning these agentic demos are rapid advancements in code-language reasoning, agentic trajectory planning (approaches vary across prompting, behavioral cloning / fine-tuning, reinforcement learning), and various agent-computer interface (ACI) improvements (i.e., abstractions and infrastructure across the browser and operating system that enable seamless agentic tool querying and self-correction).

3. Code-language reasoning will remain an epicenter of AI activity, set up to benefit from both model layer innovation (e.g., GPT-4, Claude 3 Opus) and novel reasoning/agentic paradigms (e.g., Cognition’s Devin, SWE-agent, OpenDevin). Model layer improvements will flow down into code editing and completion quality, with value ultimately accruing to developers and software organizations. Beyond code reasoning, systems that push the boundaries of latency, context size, and expand the language domain / pre-training set will also drive outsized value for developers.

AI is driving both innovation and upheaval alike, and accelerating developer velocity, productivity, and leverage for software organizations. Forward-thinking software organizations are routinely surveying the landscape for emerging tools and vendors, and rapidly prioritizing and adopting high-value developer software. Developer budgets are flowing once again and the willingness to pay is high for tools that have visible impact.

For developer entrepreneurs, this is an exciting time to be building; opportunities abound across copilots but also infrastructure, dev tooling, QA, IT configuration and provisioning, security operations monitoring, penetration testing, and on and on.

Copilots are perhaps the most obvious opportunity at the moment, but that makes them likely the most competitive playing field. We have also seen an explosion of tools in more specific developer domains — from SecOps in security to SRE to QA and pen testing. These tools use LLMs to abstract away low-level complexity and automate highly time-consuming, painful engineering tasks, freeing up engineering resources for higher-order tasks. The integration of AI in DevOps processes will enhance CI/CD pipelines, automated testing, and deployment strategies, leading to faster and more reliable software delivery.

Code refactoring is another great example of AI’s impact in the developer workflow and ecosystem. Many modern engineering teams spend only a fraction of their FTE time writing net-new code. At large organizations in particular, a large fraction of SWE time is spent on the less “sexy” parts of the software engineering role: maintaining, securing, and testing code. Many of these tasks, such as code refactoring, require deep knowledge of the stack and are often unwieldy projects performed with dread by senior engineers.

AI has obvious potential to address these challenges; startups like Gitar, Grit, ModelCode, and others leverage code-gen models, static analysis, and AST parsers to interpret code structure and migrate code across language, package libraries, and frameworks. Some of these efforts are focused on modern web frameworks while others work with brittle legacy engineering stacks (i.e., COBALT, PEARL, etc.) where fluent engineers are becoming obsolete over time. Many workflows adjacent to the core software engineering function are also highly time-intensive, repetitive, and ripe for automation.

Prediction: By 2030, a majority of corporate software developers will become something more akin to software reviewers. The cost of development will fall and as experienced developers become more productive their salaries will rise.

AI will impact the scope, and skills required for all job markets, but perhaps none more so than of the developer. AI enhancements will not only vastly improve the productivity of this occupation, but also expand the boundaries of the developer universe. By the end of this decade, development capability will be an accessible skill to most of the global population.

Deep Dive: A behind-the-scenes conversation on AI Developer Tools 

Trend 3: Multimodal models and AI agents will transform human relationships with software

The rise of multimodal models and AI Agents is leading the next wave of innovation in AI, and dramatically expanding AI’s potential applications to far broader use cases than early text-based models achieved. There’s a greenfield opportunity for AI entrepreneurs to innovate across new modalities, such as voice, image and video, as well as agentic workflows. These new modalities give AI the equivalent to the human capabilities of vision, hearing and speech, which unlocks the opportunities for AI to play a role in augmenting the large share of human work that is dependent on these senses.

In the next 12 months, we expect voice AI applications in particular to see breakout growth. Over the longer term we also see the promise of agent-first products changing the way businesses operate, as they set new expectations in terms of the complexity and breadth of tasks that AI can be entrusted to handle.

Prediction: Voice AI applications will unlock $10B of new software TAM over the next five years.

Recent progress is undeniable:

Voice

The first wave of voice AI companies were primarily leveraging advancements in Automatic Speech Recognition (ASR), such as Abridge, which offers the leading product for transcribing notes from doctor-patient conversations, and Rillavoice, which captures field sales representative conversations with customers to assist in sales training. We are now seeing a new wave of voice AI companies that are developing conversational voice products capable of handling tedious and repetitive workflows, empowering humans across sales, recruiting, customer success, and administrative use cases to focus on higher-value work. One example from our portfolio is Ada, which has taken advantage of recent voice breakthroughs to expand their chat-based customer support product to incorporate conversational voice.

Underpinning these developments are new voice architectures. We are seeing a shift from cascading architectures (ASR transcribes audio to text which is passed into an LLM, then text is fed back into a Text-to-Speech model) to speech-native architectures, as exhibited by new models such as GPT-4o, that can process and reason on raw audio data without ever transcribing to text and respond in native audio. This transition will enable conversational voice products with much lower latency and much greater understanding of non-textual information like emotion, tone, and sentiment, most of which get lost in cascading architectures. These advancements will result in conversational voice experiences that are truly real-time and can help users resolve their issues faster and with far less frustration than prior generations of voice automation.

AI voice applications are emerging in many industries including auto dealerships, retail, restaurants, and home services. A large portion — or even a majority — of inbound sales calls are missed as they happen outside of business hours, and in these cases, AI is primed to pick up the slack. AI voice applications in sales tend to be incredibly high ROI use cases because the AI is essentially picking up lost revenue for these businesses, and can thus offer a really compelling value proposition.

Entrepreneurs building at the forefront of voice AI are more equipped than ever to deliver interfaces that are increasingly natural and conversational, capable of providing near human-level performance. We expect to see an explosion of companies across the Voice AI stack (see below), many of which will experience truly breakout growth. In the process, we also expect consumer expectations around interacting with voice AI to change, as modern conversational voice applications start to deliver far more natural experiences for users and ultimately get them to resolution much faster.

Image / Video

Computer vision models have existed for years, but what is so exciting about the new generation of multi-modal LLMs is their ability to combine their understanding of image and text data (among other modalities), as this combination is extremely useful for many tasks.

The initial wave of enterprise-based image applications was focused largely on data extraction use cases. We have seen companies like Raft ingest freight documents, extracting critical information to populate the customer’s ERP and automate invoice reconciliation workflows. As the underlying models keep improving, we believe we will see a host of vertical-specific image and video processing applications emerge that will also be able to ingest increasing amounts of data to power their applications.

We have also seen applications in engineering and design that leverage vision models, and image generation models to help reason on graphical data, like schematics, or generate renderings of a building design. For example, Flux.ai offers an AI copilot that helps electrical engineers generate printed circuit board components in their design software, based on ingesting a PDF spec sheet for the component.

Autonomous AI Agents

One of the most exciting emerging themes in AI is the development of AI Agents, capable of handling complex multi-step tasks end-to-end, fully autonomously. While most AI agents don’t yet operate reliably enough to function autonomously in complex use cases, progress on agentic workflows is moving very quickly and we are seeing glimpses of what is possible. Each new demo is better than the last, with Cognition AI’s Devin — the AI software engineer — hinting at what’s possible as AI’s planning and reasoning capabilities continue to expand.

More applications are beginning to implement AI agents in highly constrained use cases in which they can limit the impact of compounding errors across multistep processes. For example, enterprises are leveraging solutions like Bessemer portfolio company Leena AI providing AI agents to support employees with IT, HR, and Finance related tasks, helping these teams free themselves of busywork and improving the employee experience.

In addition, new models are emerging with stronger reasoning capabilities that can further empower agents to execute more complex workflows. And perhaps more interesting, there is a flurry of research focused on new architectural approaches to improve agent implementations through various methods including, chain-of-thought reasoning, self-reflection, tool use, planning, and multi-agent collaboration.

2023 was the year we saw an explosion of AI applications focused on text-based use cases. In 2024, we predict multimodal models will open up new frontiers in terms of both the capabilities and use cases that we see AI being used in at the application layer. This will lead to a new wave of applications bringing near human capabilities to markets ranging from large enterprises to small businesses within specific verticals, and will even unlock exciting potential for consumer apps. 

Deep Dive: A behind-the-scenes conversation on Multimodal Models and AI Agents 

Trend 4: Vertical AI shows potential to dwarf legacy SaaS with new applications and business models

Vertical SaaS proved to be a sleeping giant that transformed industries during the first cloud revolution. Today, the top 20 US publicly traded vertical SaaS companies represent a combined market capitalization of ~$300 billion, with more than half of these companies having IPO'd in the last ten years.

Early signal on Vertical AI business model strength from the Bessemer portfolio

Bessemer was fortunate to back the legacy SaaS leader in several verticals — and now we have one of the largest Vertical AI portfolios, particularly with businesses that have reached mid to growth stages. As a result, we already have meaningful data we can use to compare Vertical AI companies and legacy vertical SaaS comps. And while we’re as excited as any VC about the power of language models, we’re growing equally excited by the early data we’re seeing on Vertical AI business models. Three analyses of our Vertical AI portfolio hint at the strength of this new class of applications.

Deep Dive: A behind-the-scenes conversation on Vertical AI

Trend 5: AI brings Consumer Cloud back from the dead

It’s no secret that consumer cloud has had a slow decade. (We define consumer cloud as companies that provide cloud-based storage, compute, and digital applications directly to individual consumers, including, at times, concurrent B2B and “prosumer” offerings.) 

Deep Dive: A behind-the-scenes conversation on Consumer Cloud