Google alters its cloud playbook, enlisting partners to sell more AI chips

Alphabet said Oracle will offer Google TPUs, making them available to customers outside Google’s infrastructure for the first time and expanding its effort to compete in AI chips dominated by Nvidia. Analysts’ TPU revenue estimates vary widely: TD Cowen models $9.4 billion in second-half sales and $62.5 billion in 2027, while Wells Fargo expects about $9 billion in third-quarter sales and $99 billion next year; Google CFO Anat Ashkenazi said the vast majority of revenue from existing agreements would be recognized in 2027. External TPU sales could carry an estimated 20% operating margin, below Google Cloud’s roughly 35% margin, while Anthropic and reportedly OpenAI are among prospective TPU customers.

Analysis

The strategic value is less “TPUs replace Nvidia” than Google monetizing capacity it cannot fully deploy internally: external access can improve utilization and broaden the software/customer ecosystem, but may dilute economics versus using the chips to differentiate Google Cloud. If external TPU sales earn materially below cloud margins, headline TPU growth could coexist with slower cloud-margin expansion; verify recognized revenue, utilization, and incremental capex rather than relying on large forward estimates.

Near term (earnings/days), expectations embedded in widely divergent TPU forecasts create asymmetric execution risk: guidance that emphasizes contracted demand without delivery capacity or revenue recognition may disappoint. Over 1–3 months, Oracle’s role could help Google reach customers that do not want to migrate workloads, while making Oracle a distribution beneficiary; it does not establish durable demand or disclose the economics of the arrangement. Over 6–18 months, heterogeneous-chip procurement by Anthropic and potentially OpenAI weakens Nvidia’s pricing power at the margin, but migration costs, software tooling, and workload-specific performance make abrupt substitution unlikely. AMD may benefit from buyers seeking a second source, though TPUs are a more direct competitive alternative for some workloads.

Contrarian view: the market may overvalue announced capacity and underweight financing and concentration risk. Anthropic’s leasing structure supports deployment but adds dependence on continued financing and future workload demand; it also lets Google sell compute to a model competitor. A failure to convert commitments into recognized, utilized revenue—or evidence of cloud-margin deterioration—would challenge the growth narrative.

Key Decisions for Investors

• Avoid chasing GOOG solely on headline TPU estimates ahead of earnings. Reassess after disclosure of TPU revenue recognition, utilization, external-customer mix, and cloud operating margin; a widening gap between bookings and recognized revenue is a warning.
• Conditional relative-value idea: favor GOOG over NVDA on a post-earnings pullback if Google confirms capacity delivery and customer utilization without material cloud-margin slippage. Falsify if cloud margins contract beyond management’s explanation, deployment slips, or TPU demand proves concentrated in a few financed buyers.
• Treat AVGO as a secondary beneficiary, not a clean TPU-volume proxy: its silicon and networking exposure could rise with deployments, but verify order conversion and customer/financing concentration before adding. Watch for leasing commitments that fail to translate into funded, operating capacity.
• Monitor ORCL for evidence that TPU access attracts incremental workloads rather than merely shifting infrastructure economics. Without customer adoption or disclosed commercial contribution, the partnership alone is not a basis for a position.