Google's 2026 AI Infrastructure: Custom TPUs and NVIDIA GPUs Strategy

Google’s 2026 AI infrastructure investment plan (custom TPUs vs. NVIDIA GPUs)

Based on Alphabet’s 2026 Q1 earnings transcript excerpts, Google is not describing a specific “2026 capex slice” that quantifies how much incremental money will go to custom TPUs versus NVIDIA GPUs. Instead, management frames 2026 as an overall compute buildout/capacity allocation plan across Google Cloud and its AI stack, where custom TPUs are a major component (including externally sold TPU hardware) and NVIDIA GPUs are also part of the first-line AI accelerator portfolio.

1) Overall 2026 AI infrastructure spending: higher CapEx, driven by AI compute demand

Alphabet updated its full-year 2026 CapEx guidance range to $180B–$190B, explicitly “to include investment related to the acquisition of Intersect” and to reflect demand for AI compute resources. ¹² Management also emphasized “unprecedented internal and external demand for AI compute resources,” and expects 2027 CapEx to “significantly increase compared to 2026.” ²³

Implication for AI infrastructure in 2026: the company is actively investing across its compute stack in 2026, with the major financial signal being the upward CapEx range rather than a TPU-vs-GPU dollar breakdown. ²³

---

2) Custom TPUs: actively scaling, with both internal use and select third-party sales

A) TPUs are part of the “full stack” compute approach and ROIC discipline

Management described using TPUs within a broader Google Cloud “full stack of AI” framework and said they apply an ROIC approach to decisions about infrastructure and (in some cases) direct TPU hardware sales. ¹⁴

B) Third-party TPU sales exist, but timing skews to later years

Alphabet disclosed that it expects to begin recognizing a “small percent” of revenues from TPU hardware agreements later in 2026, with the “vast majority” of revenue realized in 2027. ²⁵ This is consistent with the statement that revenue recognition depends on when TPUs are shipped. ²

Management also described how these TPU agreements show up in backlog:

• The TPU hardware agreements referenced earlier are reflected in Google Cloud backlog of $462B at quarter end. ⁵⁶
• “Just over half” of the total backlog is expected to convert to revenue in the next 24 months. ⁵
• Specifically for TPU hardware sales, only a small percent is expected to come through as revenue later in 2026, and the majority in 2027. ⁵

C) TPU hardware is being positioned as a differentiated accelerator (TPU generations and capabilities)

Alphabet highlighted its 8th-generation TPUs (TPU 8t and TPU 8i) and their intended roles:

• TPU 8t for high-performance model training, with performance comparisons vs prior generation materials mentioned by management. ⁷
• TPU 8i for low-latency inference, with a “performance per dollar” comparison also described by management. ⁷

Implication for “plans for 2026” regarding custom TPUs: Google’s disclosed plan for 2026 includes (i) continuing TPU-based capacity buildout as part of Cloud/AI infrastructure and (ii) selective external TPU hardware monetization, but with revenue timing largely weighted toward 2027. ¹²⁵

---

3) NVIDIA GPUs: a core part of the accelerator portfolio alongside custom silicon

Management stated that custom TPUs, Axion CPUs, and the “latest NVIDIA GPUs” form the “widest variety of compute options.” ⁸ NVIDIA GPUs were described as:

• “a core part of our AI accelerator portfolio,” and
• slated to be among the first to offer NVIDIA Vera Rubin NVL72 in addition to Blackwell and Hopper-based instances already available. ⁸

Alphabet also stated there is “tremendous demand” for both “AI infrastructure, including massive interest in our GPU offerings as well as TPUs,” indicating both are actively in consideration for capacity allocation. ⁹¹⁰

Implication for 2026 GPU plans: the excerpted disclosure suggests NVIDIA GPUs remain a standard, actively deployed compute option in 2026, with new NVIDIA platform/instance offerings (e.g., NVL72) expected early as part of the accelerator roadmap. ⁸⁹

---

Direct answer to the user question: what does Google say for 2026 about AI infrastructure investment in custom TPUs vs NVIDIA GPUs?

1. Custom TPUs (2026 plan):

   • Google treats TPUs as part of its broader Cloud/AI infrastructure strategy with an ROIC approach, and it sells TPUs both for customer deployments and selectively for third parties. ¹⁶
   • For 2026 specifically, management expects only a small percent of TPU hardware agreement revenues to be recognized later in 2026, with the vast majority in 2027—implying 2026 investment supports ramps/shipments that monetize mostly after 2026. ²⁵

2. NVIDIA GPUs (2026 plan):

   • NVIDIA GPUs are described as a core component of Google’s AI accelerator portfolio and are deployed alongside custom TPUs.
   • The company expects GPU instances to include NVIDIA Vera Rubin NVL72 availability “among the first,” in addition to existing Blackwell/Hopper-based instances—indicating active 2026 infrastructure enablement and refresh. ⁸

3. CapEx magnitude (but not a TPU-vs-GPU split):

   • Alphabet’s disclosed 2026 AI infrastructure investment signal is the overall CapEx increase to $180B–$190B (with AI compute demand described as unprecedented), rather than a quantified TPU vs GPU dollar allocation. ²³

---

Bottom line

In the provided 2026 Q1 excerpts, Alphabet indicates that its 2026 AI infrastructure investment will support both custom TPU capacity (including selective external TPU hardware sales, with revenue mostly in 2027) ²⁵ and NVIDIA GPU deployment/updates (including early availability around Vera Rubin NVL72), as part of a diversified compute portfolio. ⁸⁹ However, the excerpts do not provide a specific 2026 investment budget split between custom TPUs and NVIDIA GPUs in dollar terms; they instead describe overall 2026 AI CapEx and how TPU revenue recognition and GPU instance offerings fit into the broader roadmap. ²⁸³