Alphabet Sold Its AI Chips to Outside Data Centers for the First Time. The Cloud Line Now Holds the Revenue.

For the first time, Alphabet sold its in-house AI chips (Tensor Processing Units, or TPUs) to data centers operated by third parties, and the revenue is now reflected inside Google Cloud's segment. Alphabet has not disclosed the TPU-specific dollar figure; it is almost certainly a small slice of the $24.8 billion cloud total. What crossed in Q2 is not a dollar threshold but a unit-of-account threshold: Alphabet's AI compute is no longer reported only as an internal cost or as bundled cloud capacity. Some of it is now reported as chip-vendor sales.

The Motley Fool's re-reporting of Alphabet's Q2 earnings called TPU direct sales "the first time TPUs have generated revenue for the company." Until now, TPUs were either an internal cost (training and serving Google's own models) or bundled capacity inside Google Cloud, where customers leased access to TPU-backed compute without buying a chip. Selling TPUs to data centers operated by other companies moves the revenue line from a service sale to a chip-vendor sale. Margins look different. Revenue recognition looks different. The competitive set looks different.

The structural shift is the unit-of-account change. Alphabet is no longer just a hyperscaler that designs its own silicon. It is, in a narrow and undisclosed way, an AI chip vendor. The 10-Q filed with the SEC is the verification trail for how Alphabet classified the new revenue line, and the absence of a TPU-specific disclosure in the filing tells readers something in its own right: management chose to keep the TPU dollar figure bundled inside cloud rather than break it out as a separate chip-vendor line.

This is the hyperscaler-as-chip-vendor pattern that Amazon and Microsoft are also running. AWS ships Trainium and Inferentia. Microsoft ships Maia. Both partner with Broadcom on parts of the design. The pattern exists because vertical integration pays: a hyperscaler that designs its own chips saves the margin it would otherwise pay to Nvidia, and a chip that a hyperscaler can also sell to outside buyers recoups a piece of that design cost. Alphabet's TPU direct sales are the first time this pattern has shown up in Alphabet's reported revenue.

The competitive ceiling is Nvidia. Nvidia still designs the leading AI processors, and its data-center revenue and CUDA software moat remain the standard against which TPU, Trainium, and Maia sales are measured. Alphabet's chip-vendor lane is narrow by comparison. The lane did not exist in Alphabet's reported revenue before Q2. It is open now.

The market all of these chip vendors are chasing is large. Technavio projects worldwide AI chip revenue to grow at more than 24% annually through 2030, adding roughly $155 billion. Global Market Insights puts annual AI chip sales at roughly $1.1 trillion by 2035. Both figures are third-party projections and treated as background, not as a forecast of how much of the total Alphabet will capture.

Three signals in the next two quarters would tell readers more about the durability of the new lane than the undisclosed Q2 dollar figure. The first is whether Alphabet begins to break out a TPU contribution in the next 10-Q, or whether it stays bundled inside cloud. The second is whether Nvidia addresses TPU direct sales in its next earnings call, either as a competitive footnote or as a demand-side note from its hyperscaler customers. The third is whether Amazon or Microsoft report a similar direct-sale line for Trainium or Maia. Any of those three would shift the read on whether Q2 was a one-quarter threshold or the start of a recurring chip-vendor line.