Goldman Sachs: Rampart Model Reveals AI Computing Power Shifting to Edge Devices
nashnova research
Goldman's trading-desk head Rich Privorotsky flagged a 14.7 MB browser-native privacy model called Rampart as a sign that AI compute is migrating from the cloud to local devices — a shift that challenges the market's linear extrapolation of hyperscale demand.
What is Rampart, and why did a Goldman trading desk highlight it?
Rampart is a PII-redaction model — a tool that automatically detects and masks personal data such as names and ID numbers — built by National Design Studio. It is just 14.7 MB, runs entirely in the browser via WebGPU, and sends no data to any server.
Performance: 98.4% recall on a holdout set, with latency in the low milliseconds. This means → a tiny model can now do locally what used to require a cloud-hosted frontier model, and do it fast.
Goldman's One-Delta desk head Rich Privorotsky stressed in his morning note that the point is not Rampart itself — it is the direction: more and more AI tasks are moving from the cloud back to the device.
What exactly is migrating to the edge?
Privorotsky listed tasks that can already run locally: OCR, PII redaction, embeddings, retrieval, routing, intent detection, memory, and small-context summarization. All of these were assumed to need cloud frontier models a year ago.
In plain terms = twelve months ago, the market assumed nearly every AI workload required hyperscale data centers. That assumption is breaking down — a large share of "mid-to-low-tier" tasks can run on phones, browsers, or local servers.
The most critical capability here is routing — the mechanism that decides which model should handle a given request. As routing improves, calls to the most expensive frontier models drop. The frontier model increasingly becomes a "high-end reasoning layer," not the entire AI stack.
What does this mean for cloud-giant valuations?
Privorotsky posed the key question: if the market is willing to extrapolate every new AI workload into hyperscale cloud demand, it should equally factor growing edge intelligence into its valuation framework.
Put simply = the question is no longer "cloud or local" — it is "how many workflows actually reach a cloud model?" That share may be lower than current pricing implies.
This reflects a forming valuation tension: the faster edge capabilities expand, the more the growth slope of hyperscale cloud demand needs recalibrating.
Where do the mega-cap tech stocks stand right now?
On technicals, Google reclaimed its 100-day moving average, Amazon held its 200-day moving average, both bouncing sharply from oversold levels. Microsoft rallied from its year-to-date low, but Privorotsky sees its outlook hinging more on rationalizing capex against free cash flow.
Apollo chief economist Torsten Slok pointed to the deeper issue: "There is no sign of margin improvement outside the tech sector."
This means → today's AI valuations rest entirely on a promise — that margins at the other 493 companies in the S&P 500 will eventually rise because of AI. That promise has yet to show up in the data, and the coming earnings season is the test window.
市场有风险，内容仅供研究参考，不构成投资建议。