Generative AI saw faster and more widespread adoption in 2024 than any other technology category, with many companies seeing ROI and scaling the use cases into wider usage. The cost of machine intelligence has fallen by 1000x in just three years – from $60 per million tokens with GPT-3 in 2021 to $0.06 with Meta’s Llama 3.2. Vendors added Gen AI across the board to Enterprise software products. We saw the emergence of agentic AI, multi-modal AI, reasoning AI, and open-source AI projects that rival those of the biggest commercial vendors. As we approach 2025, the world of artificial intelligence (AI) is set to undergo monumental shifts across industries, from healthcare to automation, transportation, and beyond. AI technologies are becoming increasingly sophisticated, enabling new applications, improving operational efficiencies, and reshaping how businesses and societies operate. Below are key trends and technologies expected to dominate the AI landscape in 2025.
1. Generative AI Models will Become Commodities
The generative AI landscape is evolving rapidly, with foundation models seemingly now a dime a dozen. As 2025 begins, the competitive edge is moving away from which company has the best model to which businesses excel at fine-tuning pretrained models or developing specialized tools to layer on top of them. The boom in generative AI models is similar to the PC industry of the late 1980s and 1990s. In that era, performance comparisons focused on incremental improvements in specs like CPU speed or memory, similar to how today's generative AI models are evaluated on niche technical benchmarks. Over time, however, these speed and feed distinctions faded as the market reached a good-enough baseline, with differentiation shifting to factors such as cost, UX and ease of integration. Foundation models seem to be on a similar trajectory: As performance converges, advanced models are becoming more or less interchangeable for many use cases. In a commoditized model landscape, the focus is no longer number of parameters or slightly better performance on a certain benchmark, but instead usability, trust and interoperability with legacy systems. The end-to-end AI system will matter more than the models. An AI model in isolation is just bits on a disk. While the last four years were defined by the race for scale, 2025 will be shaped by researchers and builders who master this systems-level architecture. The breakthroughs we’ll see won’t come simply from training larger models, but from finding more elegant and effective ways to combine multiple, smaller models and software components. This move from “model-centric” to “system-centric” thinking will start to erode incumbents’ capital advantages and benefit startups who can move quickly and experiment. In that environment, AI companies with carefully architected ecosystems, user-friendly tools and competitive pricing are likely to take the lead.
2. OpenAI’s First-Mover Advantage will Erode
The history of innovation is filled with pioneering stories of Yahoo, Netscape, MySpace that catalyzed technology revolutions but failed to capture their value. We think in 2025, OpenAI, despite the impressive technical achievements of AI models like o3 and its recent $150B valuation may join their ranks as the foundational model segment becomes more competitive. For example, Google’s Gemini has already surpassed GPT-4 on key industry benchmarks, while Meta’s open-source strategy delivers comparable capabilities at half the cost. When Llama 3 powers free AI features across Facebook, Instagram, and WhatsApp that reach 4B users, ChatGPT’s 10M paying users starts to look less like market dominance and more like a craze driven early lead. Open-source progress is equally striking – these models now match their closed counterparts on nearly every benchmark that matters. To give just one data point, Llama 3.1 405B sits just a hair behind Claude 3.5 Sonnet and GPT-4 Turbo on MMLU. Enterprise spending patterns back up the evals. Data from Ramp shows that OpenAI’s share of AI spend among customers on their platform has dropped from 90% to 76% this year. Businesses are adopting a multi-model strategy and building infrastructure to easily switch between providers. Excellence in model development alone, it turns out, doesn’t create customer lock-in.
3. AI Apps and Data Sets will Become More Use Case Specific
Leading AI labs, like OpenAI and Anthropic, claim to be pursuing the ambitious goal of creating Artificial General Intelligence (AGI), commonly defined as AI that can perform any task a human can. But AGI -- or even the comparatively limited capabilities of today's foundation models -- is far from necessary for most business applications. For enterprises, interest in narrow, highly customized models started almost as soon as the Generative AI hype cycle began. A narrowly tailored business application simply doesn't require the degree of versatility necessary for a consumer-facing chatbot. Although, historically, larger data sets have driven model performance improvements, researchers and practitioners are debating whether this trend can hold. Some have suggested that, for certain tasks and populations, model performance plateaus -- or even worsens -- as algorithms are fed more data. The motivation for scraping ever-larger data sets may be based on fundamentally flawed assumptions about model performance. That is, models may not, in fact, continue to improve as the data sets get larger -- at least not for all people or communities impacted by those models.
4. Meta's Llama could Become Linux of AI World
Meta’s Llama architecture may become for AI what Linux became for data center servers – the standard that defines how AI systems are built and deployed. Building on Llama enables developers to tap into an entire ecosystem that’s being optimized around it, from hardware and dev tools to training and deployment pipelines. This marks a break with the AI development paradigm of 2022-2023, when entering the model space meant raising hundreds of millions just for initial training runs. The question wasn’t “What innovative approach can we take?” but “Can we access a 100,000-GPU cluster?” In 2025, with continued advances in open source and model distillation, small teams will increasingly compete against the bigger labs – particularly for specific vertical and “last mile” use cases where specialized knowledge matters more than scale.
5. Small Language Models for Edge Computing
Most of the attention in 2024 was on the big language models specifically on ChatGPT in its various permutations, as well as competitors like Anthropic’s Claude and Meta’s Llama models. But for many business use cases, LLMs are overkill and are too expensive, and too slow, for practical use. Looking ahead to 2025, we expect small language models, specifically custom models, to become a more common solution for many businesses. LLMs aren’t just expensive, they’re also very broad, and not always relevant to specific industry use cases.
Smaller models, on the other hand, are more tailored, allowing businesses to create AI systems that are precise, efficient, robust, and built around their unique needs. They can be more easily trained on a company’s own data. Small language models are also better for edge and mobile deployments, as with Apple’s recent mobile AI announcements. As the demand for faster processing and lower latency increases, edge AI will see rapid growth. Edge AI refers to running AI algorithms directly on local devices (such as smartphones, IoT devices, or drones), rather than relying solely on centralized cloud servers. This trend will enable real-time decision-making in various industries, including agriculture (smart farming), manufacturing (predictive maintenance), and retail (inventory management and customer personalization).
The combination of Edge AI and IoT will lead to more intelligent environments, where devices communicate with each other and act autonomously. This integration will make smart homes, factories, and cities more efficient, with AI making data-driven decisions to improve user experiences, reduce energy consumption, and enhance safety. Enterprises, especially those with large employee and customer bases, will set the standard for on-device AI adoption. And we’re likely to see an increase of tech providers keeping large enterprises top of mind when developing the on-device technologies.
6. NVIDIA’s Near-Monopoly on AI Hardware will Erode
The story of AI hardware in 2024 was largely the story of NVIDIA – their near-monopoly on AI microchips/GPUs drove the company to a $3.3T valuation. But we are seeing mounting competition and a shift in how AI systems consume computational resources. AI model pretraining cemented NVIDIA’s dominance – as it requires massive clusters of chips running at full capacity for months, processing enormous batches of data in parallel. NVIDIA with its GPUs excelled by building an integrated stack of hardware and software optimized for these concentrated, predictable workloads. Yet AI inference presents a different set of challenges: workloads are spiky and unpredictable, latency matters more than raw throughput, and computation needs to happen at the edge, rather than in centralized data centers. So, the AI infrastructure landscape of 2025 will likely become more distributed and heterogeneous, optimized for different tradeoffs than today’s massive GPU farms. There’s a significant opening for competitors – both from custom silicon designed by tech giants (Apple, AMD, Microsoft, Meta, Google, Amazon, and Tesla are all contenders) and from innovative startups. The question isn’t whether NVIDIA remains a major player, but whether they can maintain their near-monopolistic position.
7. AI-related Security Concerns will Escalate
The widespread availability of Generative AI, often at low or no cost, gives threat actors unprecedented access to tools for facilitating cyberattacks. That risk is poised to increase in 2025 as multimodal models become more sophisticated and readily accessible. In a recent public warning, the FBI described several ways cybercriminals are using generative AI for phishing scams and financial fraud. For example, an attacker targeting victims via a deceptive social media profile might write convincing bio text and direct messages with an LLM, while using AI-generated fake photos to lend credibility to the false identity. AI video and audio pose a growing threat, too. Historically, models have been limited by telltale signs of inauthenticity, like robotic-sounding voices or lagging, glitchy video. While today's versions aren't perfect, they're significantly better, especially if an anxious or time-pressured victim isn't looking or listening too closely. Audio generators can enable hackers to impersonate a victim's trusted contacts, such as a spouse or colleague. Video generation has so far been less common, as it's more expensive and offers more opportunities for error. But, in a highly publicized incident earlier this year, scammers successfully impersonated a company's CFO and other staff members on a video call using deepfakes, leading a finance worker to send $25 million to fraudulent accounts. Other security risks are tied to vulnerabilities within models themselves, rather than social engineering. Adversarial machine learning and data poisoning, where inputs and training data are intentionally designed to mislead or corrupt models, can damage AI systems themselves.
8. The Limits of Model Pretraining will Drive new AI Breakthroughs
The assumed progress of scaled pretraining has hit data wall, energy wall, and model architectures. But we think in 2025, these won’t limit AI’s advance – they’ll redirect it toward new AI approaches. One of the promising frontiers is “reasoning” – where models don’t just recall patterns from training but actively work through problems during inference. For example, OpenAI’s o3 model: rather than producing instant answers, it generates detailed reasoning paths tailored to each task. As a result, o3 has achieved 87.5% on the ARC-AGI prize and 25% on FrontierMath (a specialized math test written by Fields Medalists, where previous models peaked at 2%). To put this leap in perspective: it took four years for performance on ARC-AGI to inch from 0% with GPT-3 to 5% with GPT-4. o3 represents a fundamental breakthrough in AI’s capacity to handle novel situations. But this type of inference strategy comes at a cost: o3’s top-performing version demands 172x more compute than its baseline, costing over $3,400 per answer. But if the past three years have taught us anything, it’s that these costs tend to plummet. The convergence of more efficient training and sophisticated reasoning suggests that AI progress in 2025 may accelerate beyond even 2024 impressive gains.
9. Multi Modal AI will Become Critical
Humans are multi-modal - we read and write text, we speak and listen, we see, and we draw. And we do all these things through time, so we understand that some things come before other things. Today’s AI models are, for the most part, fragmentary. One can create images, another can only handle text, and some recent ones can understand or produce video. When people want to do speech generation, they go to a specialized model that does text to speech. Or a specialized model for image generation. To have a full understanding of how the world works, for true general intelligence, AI has to function across all the different modalities. Some of this is available today, though usually the multi-modality is an illusion, and the actual work is handled behind the scenes by different specialized, single-mode models. Architecturally, these models are separate, and the vendor is using a mixture-of-experts architecture. In 2025, however, we expects multi-modality to be an important trend. Multi-modal AI can be more accurate and more resilient to noise and missing data and can enhance human-computer interaction. Gartner, in fact, predicts that 40% of gen AI solutions will be multi-modal by 2027, up from 1% in 2023.
10. Agents will Replace Services
Software has evolved from big, monolithic systems running on mainframes, to desktop apps, to distributed, service-based architectures, web applications, and mobile apps. Now, we think it will evolve again. Agents are the next phase. Agents can be more loosely coupled than services, making these architectures more flexible, resilient and smart. And that will bring with it a completely new stack of tools and development processes. Today, AI agents are relatively expensive, and inference costs can add up quickly for companies looking to deploy massive systems. But that’s going to shift. And as this gets less expensive, the use cases will explode. In addition to agents replacing software components, we’ll also see the rise of agentic assistants. For example, the task of keeping up with regulations. Today, consultants get continuing education to stay abreast of new laws or reach out to colleagues who are already experts in them. It takes time for the new knowledge to disseminate and be fully absorbed by employees. But an AI agent can be instantly updated to ensure that all our work is compliant with the new laws. Soon, showing up to a meeting without an AI assistant will be like an accountant trying to do their work without Excel. If you’re not using the proper tools, that’s your first indication you aren’t the right person for the job.
The next step for agents is pulling together communications from all the different channels, including email, chat, texts, social media, and more. Making better spreadsheets doesn’t make for great headlines, but the reality is that productivity gains from workplace AI agents can have a bigger impact than some of the more headline-grabbing AI applications. Things are going to get really interesting when agents start talking to each other. It won’t happen overnight, of course, and companies will need to be careful that these agentic systems don’t go off the rails. First, an agent has to be able to recognize whether it’s capable of carrying out a task, and whether a task is within its purview. Today’s AIs often fail in this regard, but companies can build guardrails, supplemented with human oversight, to ensure agents only do what they’re allowed to do, and only when they can do it well. Second, companies will need systems in place to monitor the execution of those tasks, so they stay within legal and ethical boundaries. Third, companies will need to be able to measure how confident the agents are in their performance, so that other systems, or humans, can be brought in when confidence is low. If it goes through all of those gates, only then do you let the agent do it autonomously. Companies such as Sailes and Salesforce are already developing multi-agent workflows. Combine this with chain-of-thought reasoning, or the ability for an AI agent to reason through a problem in multiple steps (recently incorporated into the new ChatGPT-o1 model), and we’ll likely see the rise of domain expert AI that’s available to everyone.
11. AI-native Startups will Challenge Incumbent SW Vendors
2025 will witness the rise of a new generation of enterprise software vendors. These won’t be traditional systems with AI features bolted on, but AI-native platforms that reimagine how software works. Consider what’s happening in CRM, traditionally one of enterprise software’s most entrenched markets. Today’s systems of record – Salesforce, Hubspot, etc. – were built around structured representations of data in text-based formats. An AI-native sales platform doesn’t just add features to this aging model: it reimagines the core system as a multimodal brain that processes and acts on text, image, voice, and video. Incumbents’ distribution moats – often an insurmountable barrier for startups – could matter less when the underlying technology difference is this profound. Sales teams aren’t adopting AI-native platforms because they’re incrementally better, but because they eliminate entire work-flows – from lead research to call preparation to collateral creation. While cloud and mobile each produced about twenty startups with $1B+ revenues, those companies had to find narrow vertical niches to compete. AI’s advances now enable startups to launch frontal attacks on nearly every major category of enterprise software – from sales and marketing automation to ERP and financial planning. Any form of legacy software that’s anchored on a structured-data, text-dominant paradigm could become obsolete. The opportunity to rebuild these products from the ground up with AI-native architectures represents one of the largest value creation opportunities in enterprise software history.
12. Mass Customization of Enterprise Software
Today, only the largest companies, with the deepest pockets, get to have custom software developed specifically for them. It’s just not economically feasible to build large systems for small use cases. People are all using the same version of Teams or Slack. Microsoft can’t make a custom version just for me. But once AI begins to accelerate the speed of software development while reducing costs, it starts to become much more feasible. Imagine an agent watching you work for a couple of weeks and designing a custom desktop just for you. Companies build custom software all the time, but now AI is making this accessible to everyone. We’re going to start seeing it. Having the ability to get custom software made for me without having to hire someone to do it is awesome.
13. Robotaxis will Get Public Confidence
In 2025, self-driving cars will transform public trust in AI by making machine intelligence visible and undeniable. While we can debate a chatbot’s abilities, there’s no arguing with an AI that navigates the road more safely than humans. This every day, physical proof of AI’s societal benefits will do more to win public confidence in this technology than any model breakthrough. As of mid-2024, Waymo’s robotaxis have logged over 22 million miles of autonomous driving, including 5.9 million in San Francisco, where their white Jaguar SUVs have become fixtures of the urban landscape. Seeing our driverless future continue to come to fruition – with all its positive implications for safety, accessibility, urban design, human productivity, and overall quality of life – is one of the developments we are looking forward to in 2025.
14. AI will Change Economics SW Consumption
2025 will see AI companies will target the vastly larger services market – a roughly 10x expansion in TAM. They’ll succeed by selling actual work completion rather than just workflow enablement. This shift to outcomes-based pricing presents a classic innovator’s dilemma for incumbents. Their revenue models, sales incentives, and GTM strategies are optimized for selling seats and licenses. This opens a significant opportunity for startups building business models native to AI’s capabilities. AI is upending software’s core assumption that marginal costs approach zero at scale. Currently, each step toward higher model performance demands exponentially more resources. A chatbot that’s right 90% of the time might cost $10 per user, but achieving 99.9% accuracy could justify $1,000 per user given the underlying compute costs. We’re already seeing this pricing structure emerge with OpenAI’s latest tiers, which reach $200/month for its pro plan, with discussions of a $2000/month tier for business users. While these figures might seem steep compared to the initial $20/month offering, they’re modest when measured against human expertise. Looking ahead, as models like o3 push toward extended reasoning times of “hours, days, even weeks,” the subscription model itself may become obsolete – creating another advantage for AI-native startups over incumbents wedded to traditional pricing models.
15. Google Search will Struggle
Google’s famous “10 blue links” have defined how we access information online and shaped the architecture of the modern web. But in 2025, we’ll witness the beginning of the end for this decades-old paradigm, as AI-native information access renders traditional search results obsolete. The shift is already underway. Platforms like Perplexity and ChatGPT (which recently added web search) demonstrate how direct, synthesized answers are superior to scrolling through ad-laden link lists. More importantly, they’re training a new generation of internet users who instinctively “chat” rather than “Google” their questions. Meta’s potential entry into search could hasten Google’s fall. Meta's social graph offers something Google’s index lacks: real-time understanding of how information flows through human networks. By one estimate, Meta has access to 100x more data than what’s on the public internet – provided they can navigate the complex compliance requirements of using it. A Meta search product could combine traditional web content, social signals, and AI synthesis in ways that make Google’s current offering feel static and disconnected. Google’s ad revenue is the economic engine that built the modern web. But protecting this revenue while simultaneously shipping its replacement may prove impossible. The DOJ’s antitrust scrutiny, including discussions of forced search index licensing, further complicates Google’s ability to leverage existing advantages in this new AI era.
AI technologies in 2025 will not only redefine industries but also shape society, impacting everything from healthcare and transportation to education and entertainment. With breakthroughs in generative AI, autonomous systems, edge computing, and AI-powered healthcare, the way businesses operate, and individuals interact with technology will change profoundly. However, as these technologies evolve, ethical considerations and regulatory measures will play a key role in ensuring that AI develops in a way that benefits society as a whole, responsibly and sustainably. As AI continues to advance, it will open up new frontiers for innovation, efficiency, and creativity, making 2025 a landmark year in the history of artificial intelligence.
/Service Ventures Team
留言