Goldman Sachs: Recent AI Sector Selloff Looks More Like an Opportunity Than a Bubble Burst
nashnova research
After intensive conversations across the AI ecosystem, Goldman Sachs concludes the past four trading days' sell-off resembles a cyclical shakeout, not a bubble bursting — funding-gap fears are overstated, enterprise AI is entering production, and both threads point toward a window for valuation repair.
Why does Goldman say this isn't a bubble bursting?
Goldman compares this sell-off to the late-January-to-early-February 2026 software-stock correction — a period dominated by hard-to-disprove bearish narratives that ultimately proved to be a temporary mis-pricing.
This means → Goldman's core logic: new information is correcting overly pessimistic assumptions, not confirming them.
In plain terms = last time everyone said "this time is different" too, and it turned out to be a scare; Goldman sees the same script playing out.
How big is the "funding gap" the market fears?
Goldman believes the concern is overstated. Vendor GPU financing and customer prepayments can significantly reduce companies' own capital burden.
Non-traditional funding pools are massive: a $19 trillion private-capital market (including $4.5 trillion in dry powder, per Preqin), Nvidia's $500 billion financing facility, plus the USD investment-grade bond market with room for another $500 billion in new issuance.
This means → even if AI infrastructure spending keeps ballooning, the supply of capital is nowhere near exhausted — the "can't raise money" narrative doesn't hold up.
Why did Oracle's numbers catch the market off guard?
Oracle committed to raising only $20 billion in additional capital through 2030 ($10 billion debt + $10 billion equity) — far below Goldman's prior estimate of ~$65 billion and the Street's ~$40 billion.
The trick: three ways to get others to pay — vendor GPU financing, customer-supplied hardware, and customer prepayments — shifting roughly $110 billion of its $220 billion total capex to external parties, cutting its own burden by about 50%.
Oracle expects to turn free-cash-flow positive in H2 2028 and exceed $30 billion in FCF by FY2029 — Goldman had forecast just $3 billion; the Street expected negative $1 billion. This means → the expectation gap is enormous; if delivered, it rewrites the valuation math entirely.
How does the economics of a 1 GW data center actually work?
Oracle's illustrative framework: a representative 1 GW deployment requires roughly $37 billion in capex and, at full utilization, generates about $16 billion in annualized revenue.
Gross margin sits at roughly 35%, project-level operating margin at about 34%, supporting returns in the high-20% range.
Management says virtually all large AI infrastructure contracts use a cost-plus structure. In plain terms = hardware price swings don't eat into margins — cost changes pass straight through to the customer, locking in project profitability.
What signals are Microsoft and Google sending?
Microsoft launched a USD investment-grade bond offering reportedly of at least $25 billion, while emphasizing that customers are now deploying AI into production systems — agent-driven usage is backed by business budgets and tied to the value of work completed.
Microsoft's Foundry platform — used to build and govern AI applications and agents across 11,000 models — grew 300% year-over-year. This reflects a shift from experimentation to full production-stage enterprise AI deployment.
Google said its Carbon model ranked first on the AI index at launch, and expects training and release cadence to keep accelerating over the next 12 months via recursive self-improvement. Goldman reads all of this together: frontier-model scaling laws work, capturing the area beneath the frontier curve creates new value, and proprietary training combined with dynamic routing creates new value — all three can be true at once.
What should investors watch next?
CoreWeave is expected to report earnings on November 11. Goldman estimates fulfilling its current backlog requires roughly $20 billion in new debt; expanding from 1.5 GW to 4.2 GW would require about $40 billion. The key takeaway: bringing the locked-in 4.2 GW online requires no additional unsecured debt or equity.
Goldman sees the top priority for enterprises in 2027 as extending AI's impact beyond developers while amplifying the value of proprietary datasets.
Microsoft Foundry's growth trajectory will serve as the key barometer for enterprise AI transformation progress. This means → whether this product cycle delivers is the critical proof point for rebuilding current valuations.
市场有风险，内容仅供研究参考，不构成投资建议。