--- title: "a16z: 16 Trends in How Enterprises Build and Buy GenAI Services in 2024" author: "Unique Research" sourcePublication: "Unique Research (FFCAP English library)" originalPublishedAt: "2024-04-09T16:30:00+08:00" canonical: "https://ffcap.cn/en/research/20240409a06dxc00" source: "https://ffcap.cn/en/research/20240409a06dxc00" language: "en" cnSourceUrl: "https://view.inews.qq.com/a/20240409A06DXC00" --- # a16z: 16 Trends in How Enterprises Build and Buy GenAI Services in 2024 _Unique Research · Industry insight · April 9, 2024 · Source: a16z (Sarah Wang & Shangda Xu)_ Generative AI swept consumer markets in 2023 — past **$1B** in consumer spend. a16z argues 2024 enterprise revenue opportunity is a multiple of that. After talking with dozens of Fortune 500 / top enterprise leaders and surveying **70+**, a16z finds that in the past six months enterprises nearly **tripled** GenAI budgets, put more apps on smaller open-source models, and moved more work from experiments into production. Below: 16 considerations across resources, models, and applications. ## Resources — budgets surge, and the shift sticks ### 1\. GenAI budgets jump 2023 average enterprise spend on foundation-model APIs, self-hosting, and fine-tuning was about **$7M**. Nearly every enterprise a16z spoke with saw promising early experiments and plans **2×–5×** spend in 2024 to cover more business scenes. ### 2\. Leaders reallocate AI spend into recurring software budgets Most 2023 GenAI spend came from “innovation” pots and one-off pools. In 2024 many leaders are moving it onto permanent software budget lines; fewer than a quarter say this year’s GenAI spend will still come from innovation budgets. At the margin, some also fund GenAI from labor-cost savings — especially customer service. One firm cited ~**$6** saved per LLM-powered support call and ~**90%** overall cost reduction as a reason to raise GenAI investment **8×**. ### 3\. ROI measurement is still art and science Leaders mostly track productivity gains, using NPS / CSAT as proxies while hunting harder metrics — revenue lift, savings, efficiency, accuracy — by use case. Over the next 2–3 years ROI scrutiny will tighten; for now many still take employees’ word that they use time better, rather than hard instrumentation. ### 4\. Implementation talent is scarce — and expensive An API alone does not scale GenAI. Specialized talent for infra, maintenance, and expansion is scarce; in 2023 implementation was often the largest (or near-largest) share of AI spend. One exec: “The LLM may be only a quarter of the cost of building the scene” — development dominates. Foundation-model vendors’ professional services (often tied to custom model work) were a meaningful 2023 revenue line and a buying reason beyond raw performance. Startups that make in-house GenAI development easier should see faster adoption. ## Models — multi-model and open-source ### 5\. A multi-model future Six months ago most enterprises tested one model (usually OpenAI) or at most two. Today leaders test — and sometimes run in production — multiple models to (1) fit performance / scale / cost by app, (2) avoid lock-in, and (3) ride a fast-moving research frontier. OpenAI still dominates enterprise usage, as expected; survey charts of models under test are a leading indicator of production migration. ### 6\. Open source is booming a16z’s most surprising six-month shift: 2023 closed models held an estimated **80–90%** share (mostly OpenAI). Now **46%** of survey respondents prefer or strongly prefer open-source for 2024; in interviews nearly **60%** of AI leaders want more open-source use or would switch when tuned open models roughly match closed performance. Some firms explicitly target a **50/50** open/closed mix vs. 2023’s ~80/20. ### 7\. Control and customization beat cost as open-source reasons Control (protect proprietary data; understand why a model produced an output) and customization (effective tuning for a scene) rank well above cost. Leaders still believe GenAI’s incremental value exceeds its price — “worth paying for accurate answers.” ### 8\. Control hunger comes from sensitive apps and data security Firms remain reluctant to share proprietary data with closed-model vendors for regulatory or security reasons — especially where IP is the business. Some self-host open models; others prioritize VPC-integrated offerings. ### 9\. Leaders customize via fine-tuning / RAG, not from-scratch LLMs 2023 buzz included custom models such as BloombergGPT. In 2024 interest in customization remains, but with strong open models most firms choose RAG or fine-tuning over training an LLM from scratch. ### 10\. Cloud vendors heavily shape model-buying In 2023 many bought models through existing CSPs for security comfort and to skip long procurement — that continues in 2024. Correlation is high: Azure users lean OpenAI; Amazon users lean Anthropic or Cohere. Survey: >**72%** access models via API; over half use CSP-hosted models; >1/4 self-host (often for open models). ### 11\. Customers still chase latest leading features Beyond reasoning, reliability, and easy access (e.g. via CSP), differentiated features matter: several leaders cited Anthropic’s earlier **200K** context window; others chose Cohere for early, easy fine-tuning. ### 12\. Most enterprises see model performance converging Public benchmarks matter less than internal ones on fine-tuned open vs. closed models. Closed models often win external benches, yet leaders sometimes give open models equal or higher NPS because they fine-tune more easily. One firm: after fine-tuning, “Mistral and Llama nearly matched OpenAI at much lower cost.” ### 13\. Flexibility is the optimization target Most design apps so model swaps are an API change; some pre-test prompts for instant switchover; others run “model gardens.” Lessons from cloud lock-in plus a fast market make single-vendor bets unwise. ## Applications — more migration into real business scenes ### 14\. Build > buy (for now) Enterprises still prefer building apps in-house — few battle-tested, domain-decisive enterprise GenAI products exist (no Magic Quadrants yet). Foundation-model APIs make home-grown customer support and internal chatbots easier, while teams also try novel scenes (CPG formulas, molecular discovery, sales recommendations). GPT-wrapper startups face a dilemma: GenAI also lowers the bar for enterprises to build similar apps themselves. a16z expects outperformance from products that go beyond “LLM + UI,” rethink core workflows, or unlock proprietary data. ### 15\. Excited about internal apps; still cautious externally Two persistent concerns: hallucination / safety, and PR risk in sensitive consumer domains (healthcare, financial services). Popular 2023 apps skewed internal productivity or human-reviewed before customer touch (coding copilots, support, marketing). Fully internal apps (summarization, knowledge bots) roll out much faster than sensitive human-in-the-loop work (contract review) or customer-facing bots — firms want to avoid Air Canada–style GenAI support failures. Tools that help control those risks should see strong adoption. ## Total opportunity — large and growing fast ### 16\. Model API + fine-tuning spend >$5B ARR by end-2024 (a16z view) a16z estimates ~**$1.5–2B** ARR for model APIs (incl. fine-tuning) as of 2023 — including OpenAI via Azure. With clear enterprise signals, that slice alone should reach at least **$5B** ARR by year-end, with upside. Top-down mandates to find and deploy GenAI have compressed deals that once took a year+ into **2–3 months**, at larger ticket sizes. The opportunity extends beyond foundation models — fine-tuning tools, model serving, app builders, and AI-native applications. a16z calls this an inflection for enterprise GenAI transformation. _Original: Sarah Wang and Shangda Xu (a16z). CN edit: qiuping. Charts (12) attach on ingest._ — END — --- Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240409a06dxc00 On-site reading page: https://ffcap.cn/en/research/20240409a06dxc00