语音优先的人工智能：下一个消费者界面

Please refer to important disclosures on pages 45 and 46. Analyst certification is on page 45.William Blair or an affiliate does and seeks to do business with companies covered in its research reports. As aresult, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of thisreport. This report is not intended to provide personal investment advice. The opinions and recommendations here- Overview............................................................................................................................3Key Takeaways...................................................................................................................3AI in Existing and Purpose-Built Applications.................................................................4Revenue Models...............................................................................................................10Agentic AI Moves Beyond Simple Tasks..........................................................................16 developers and infrastructure providers have captured most of the headlines and initial invest-ments, a growing number of GenAI use-cases are emerging and will grow in importance. While theinfrastructure layer needs to be built first, as the pace of the infrastructure buildout slows, investorattention is likely to shift toward tangible applications and innovative new companies that drive that the non-infrastructure GenAI landscape is evolving rapidly, often daily. This report is our ef-fort to capture that momentum, reflect on recent developments, and highlight areas for investorsto focus on in the future. 1.AI integration is rapidly advancing as a result of greater accessibility, fueling industrygrowth.Cost reductions, open-source models, and innovations like Mixture-of-Experts haveboosted AI use in both existing applications and new start-ups. At the end of this report, we 2.AI-native companies are rewriting the revenue playbook, scaling faster, and some moreprofitably, than in prior tech cycles.Consumer-focused companies are leveraging usage-based pricing models and upselling into businesses with specific business-to-business (B2B) 3.There will soon be more AI agents than the U.S. population, and they are moving beyondsimple tasks and will transform consumer experiences.Agentic AI is evolving beyond sim-ple task execution toward autonomous, personalized workflows that could transform both en-terprise and consumer experiences over the next three to five years. We highlight the potential 4.Voice will become the main way consumers interact with AI and will unlock more AIuse-cases.In our view, voice is emerging as the primary interface for consumer AI, drivenby technical breakthroughs in speech recognition, latency reduction, and expressive text-to- 5.AI has permanently disrupted the traditional search and digital advertising landscape,although Google is likely to continue to be a winner in the space.GenAI is reshaping thesearch landscape, from the rise of AI Overviews and chat-based queries to the emergence of AIbrowsers and zero-click behavior. We explore the implications for publishers, advertisers, and William BlairAI / INSIGHTS The accessibility of AI has increased dramatically since the release of ChatGPT, lowering barriersto entry for both incumbents and start-ups. As costs fall and open-source models gain traction,there are two emerging trends: existing companies are integrating AI into legacy platforms, and Accessibility to AI Has Increased Over the Past Few Years and Is Likely to Increase FurtherSince the release of ChatGPT, the rapid pace of innovation in the AI space has increased acces-sibility to the technology, allowing more companies to compete in the evolving landscape. Thisis a result of increasing model efficiencies, competition, the emergence of open-source models,and broad enterprise and consumer adoption reducing token prices. This “democratization” ofthe technology drives costs down and allows existing companies to invest in AI using existing rev- William BlairAI / INSIGHTS The emergence ofDeepSeekin late 2024 into 2025 created a shift in the economics of the in-dustry. At the time, DeepSeek claimed that it built its V3 model, released in December 2024, forabout $6 million (a widely disputed, but not disproven, figure) using less advanced H800 chipsdeveloped by Nvidia. This called into question the massive capital expenditures by hyperscal-ers and marked a pivotal moment for AI, in our view, as it sharpened focus on cost structures offoundational models. DeepSeek was able to achieve this cost efficiency primarily through em- Although not confirmed byOpenAI,it is widely speculated that GPT-5 employs this MoE ap- proach (OpenAI mentions a “routing-first” design that resembles an expert gate network). Thiswould explain the drastic decrease in inference and token cost relative to model intelligence.While the exact parameter count of GPT-5 is