Goldman Sachs: AI Investment Focus Shifts from Computing Power to Application Monetization
nashnova research
Goldman Sachs told its Communacopia+ tech conference that global AI investment is pivoting from the compute-infrastructure phase — GPUs, memory, data centers — into a second phase where the defining question is who can turn tokens into durable cash flow.
What is the core question behind this shift?
Goldman says the thesis has changed: not "who builds the most compute" but "who converts tokens into sustainable cash flow."
This means → the market no longer rewards GPU stockpiling alone; it rewards proven, recurring revenue from AI.
Goldman names three filters: proprietary IP, innovation speed, and open architecture. Platforms with unique data assets and workflow advantages stand to be re-rated.
In plain terms = software companies with irreplaceable data win; traditional SaaS (subscription cloud software) whose features a foundation model can replicate stays under pressure.
Where do enterprises get stuck deploying AI?
Databricks CEO Ali Ghodsi said competition has moved to real-world enterprise deployment — platforms that translate model capability into reliable business outcomes will capture the next wave of growth.
Databricks' Genie product uses a proprietary indexing engine called OntoRank to organize enterprise data into a format models can compute on, acting as an "AI analyst" already used across internal sales, marketing, and finance.
In plain terms = a smart AI model alone is not enough; a company's own data must first be "translated" into language the model can read — that is the job Genie does.
As deployment scales, how is governance handled?
Goldman's analysts flag three pressures: rapid model iteration makes selection harder, rising token consumption sharpens cost and ROI scrutiny, and enterprises still lack visibility into agent data access and security risks.
Databricks addresses this with Unity Catalog. This means → the platform's commercial value lies in giving enterprises the confidence to scale deployment; stickiness comes from ongoing business value, not lock-in.
How are pricing models diverging?
Foundation-model providers still sell compute via per-token API pricing. Application-layer vendors increasingly charge by business outcome.
Examples: Intercom's Fin charges per outcome, starting at $0.99; Salesforce Agentforce offers per-action, per-conversation, and per-user-license tiers.
This means → outcome-based pricing aligns price with customer benefit, but the vendor absorbs inference, retry, and human-review costs — task success rate and per-task margin become the key commercial metrics.
Is the stock market already voting on this?
The winners: Snowflake surged roughly 17% on September 3 after reporting Q2 product revenue up 37% year-over-year to $1.49 billion, with full-year guidance raised from $5.84 billion to $6.07 billion.
This means → Snowflake's numbers prove enterprises are expanding data-platform spend for the AI application layer, and the market priced that in immediately.
The losers: on September 8, Salesforce and Intuit each fell about 4%, ServiceNow dropped roughly 5%, and the S&P 500 Software & Services index slid 1.4% — partly on fears that foundation models could commoditize traditional SaaS.
What is Goldman's bottom-line call?
The AI bull-market investment thesis is being repriced: from "who benefits from ever-rising AI capex" to "who converts compute into recurring revenue, margins, and free cash flow fastest."
In plain terms = the long-term pricing power of an AI software company hinges on whether it can plug into stronger, cheaper models while keeping its own business value irreplaceable.
市场有风险，内容仅供研究参考，不构成投资建议。