Goldman Sachs Silicon Valley Survey: AI Agents Enter the Execution Era, Compute Demand May Surge 24x in Five Years
Nashnova编辑部
Goldman Sachs wrapped its third annual Silicon Valley AI survey and reached one core conclusion: AI monetization is shifting from per-seat subscriptions to usage-, transaction-, and outcome-based pricing, with agents evolving from assistants into workflow executors — the competitive battleground is no longer who has the best model, but who controls the workflow.
What is the real barrier to enterprise AI adoption?
Not model intelligence — it is accountability, traceability, and reversibility when something goes wrong.
Stanford researchers told Goldman that most enterprises still operate in a "human-supervising-AI" mode. In law, risk management, insurance, and audit, controllability matters as much as capability.
This means → the workflows easiest to automate share three traits: clear decision boundaries, verifiable outcomes, and reversible errors. Invoice processing is the textbook case — AI extracts fields and runs checks, low-confidence cases go to a human reviewer, and booking happens through a reversible ERP process.
Frontier models vs. open-source — who wins?
Goldman's answer: not either-or, but division of labor.
The frontier camp argues that enterprise benchmarks routinely underestimate model capability; in real production, the business cost of lower accuracy can far exceed savings on inference. Several AI-native companies claim multi-model strategies, yet their core production pipelines still lean heavily on frontier models.
The open-source camp counters that most enterprise workflows do not need frontier-grade intelligence. One venture firm estimates that within roughly a year to eighteen months, about 90% of inference tokens will flow to open-source models.
In plain terms = frontier models handle high-value, high-reliability complex tasks; open-source models absorb the larger volume of standardized work and most token consumption — two lanes, each with its own business.
What are "world models," and why do they matter?
World models — models that understand physical environments, causality, and real-world dynamic interactions, rather than just processing text — are becoming a research focus. Over the past roughly eighteen months, researcher attention has visibly shifted away from large language models.
Unlike LLMs trained mainly on internet data, world models draw on physical systems, specific industries, and real operational settings. This reflects a further rise in the value of proprietary data.
This means → the problem space in physics, industry, science, and robotics is far larger than text generation alone, and the corresponding compute demand could spawn an entirely new growth curve.
How much will compute demand actually grow — and who benefits?
Goldman projects compute demand may rise roughly 24× over the next five years, with supply-demand tightness lasting longer than previously expected.
Direct beneficiaries include Microsoft, Oracle, and CoreWeave — cloud and compute-infrastructure players.
Goldman also names a verification condition: whether world models can scale on schedule is the key checkpoint for the 24× forecast to materialize.
How is AI's business model changing?
Goldman's central finding: AI monetization is shifting from per-seat subscriptions to consumption-, transaction-, and outcome-based pricing.
This means → information-service providers with trusted content, validated domain models, and established regulatory relationships are most likely to enter enterprise production environments first.
Put simply = future AI companies will not earn revenue by selling software licenses — they will bill like utilities, by usage. Whoever can embed AI into real workflows and deliver measurable results is the one who gets paid.
Content is for reference only, not financial advice.