---
title: "Google could triple TPU volumes by 2027, while doubling its debt"
description: "A GF Securities forecast puts Alphabet's tensor processing units at 8.84 million in 2027 against 2.85 million last year. Long-term debt has gone from $46.5 billion to $98.2 billion and the buyback is suspended. Nvidia reports tomorrow."
category: "Tech"
category_url: https://boursel.com/category/tech
author: "Hannah Blackwood"
published: 2026-08-25T16:37:54.000Z
updated: 2026-08-25T16:37:54.000Z
canonical: https://boursel.com/article/google-could-triple-tpu-volumes-by-2027-while-doubling-its-debt
tags: ["google", "alphabet", "tpu", "nvidia", "ai capex", "semiconductors"]
---
# Google could triple TPU volumes by 2027, while doubling its debt

A GF Securities forecast puts Alphabet's tensor processing units at 8.84 million in 2027 against 2.85 million last year. Long-term debt has gone from $46.5 billion to $98.2 billion and the buyback is suspended. Nvidia reports tomorrow.

GF Securities projects Alphabet's tensor processing unit volumes at [8.84 million units in 2027](https://finance.yahoo.com/technology/ai/articles/google-tpu-volume-could-triple-161618423.html), against 4.504 million in 2026, 2.854 million in 2025 and 2.76 million in 2024.

That is roughly a tripling in two years, and almost all of it is in the next two: units barely moved between 2024 and 2025, then rise 58 percent and 96 percent. The named generations are v7, described as Ghostfish, and the v8 family, Sunfish and Hellcat, Zebrafish and Maddog.

It is a brokerage forecast rather than company guidance, and the day before Nvidia reports is exactly when such forecasts circulate, so read it as an estimate with an audience.

## The corporate finance is the harder fact

The forecast is a projection. These are not.

Alphabet's long-term debt has risen from $46.5 billion to $98.2 billion. Second-quarter capital spending was $44.9 billion, up 100.14 percent year on year, with full-year capex guided to $175 to $185 billion. The share buyback has been suspended.

A company that borrows more than twice as much and stops returning capital to shareholders is funding something it considers more valuable than either. That is a coherent decision and it is also the exact behaviour our [morning story](/morgan-stanley-puts-the-ai-financing-gap-at-1-5-trillion-dollars-and-nobody-can) described in the aggregate, where Morgan Stanley puts $1.5 trillion of the roughly $3 trillion AI infrastructure build outside the spenders' own cash flows.

We also reported earlier that Alphabet's own commitments line [went from $322 billion to $811 billion in a single quarter](/nine-companies-have-signed-3-trillion-dollars-of-ai-commitments-that-sit-off-the). Set the three together and the picture is a company committing, borrowing and spending simultaneously, with the buyback as the release valve.

## Why the TPU number matters to somebody else's earnings

Nvidia reports after Wednesday's close, and we wrote this morning that the stock has [fallen on six of the last eight prints](/nvidia-has-fallen-after-four-straight-earnings-reports-and-it-reports-again-tomo) without missing.

TPUs are the clearest existing case of a large buyer building its own accelerators rather than purchasing them. Every unit Alphabet makes for internal use is, at the margin, a unit it does not buy, which is why volume forecasts for a chip nobody sells externally end up in stories about a company that sells chips to everybody.

Two cautions on drawing that line too hard. TPUs and Nvidia's accelerators are not interchangeable across every workload, and Alphabet remains a substantial Nvidia customer regardless. And the article supplies no comparison figures for Nvidia volumes or share, so anyone converting 8.84 million TPUs into a share-of-market claim is doing arithmetic the source does not support. We are not going to.

## The demand evidence, and its limits

Google Cloud's backlog is put at $514 billion, and Gemini is described as processing 22 billion API tokens per minute.

Backlog is contracted future revenue and is the more meaningful of the two, though it says nothing about timing or margin. Tokens per minute is a usage statistic with no price attached; it measures activity, not money, and a token processed at a loss counts the same as a profitable one.

Both are consistent with real demand. Neither establishes that the capital spending earns its return, which remains the only question that matters about any of these numbers and the one none of them answer.

*This story reports a brokerage forecast and company financial disclosures. It is not investment advice.*