Alphabet said on 28 October 2026 that Google Cloud “generates product revenues primarily from the sale of TPU systems”, and that TPU systems sales “were also a meaningful contributor to revenuegrowth” in the third quarter.
Slide 10 of its third-quarter earnings presentation shows what such a system contains, using a v7 rack as its example. On the earnings call, finance chief Anat Ashkenazi said: “A TPU system is not just a chip, but it’s a specialized system that combines multiple things, including hardware, compute, storage, networking, so we are selling much more than just a chip.”
This explainer walks through the slide’s three layers and what executives said on the call about where the systems go and how a sale turns into cash. Call quotations come from a caption transcript of the webcast.
Business Pill · Vertical Integration
An evergreen Business Engineer explainer on vertical integration, why every layer takes a cut. It explains the general idea only and does not discuss TPU systems.
The key insight: By our arithmetic from Alphabet’s slide, a TPU pod of 144 racks holds 9,216 chips if each machine carries one four-chip tray, the same count Google Cloud’s documentation lists for a TPU7x pod.
The slide lists three hardware parts. Compute is “Custom TPUs & high-performance host CPUs”. Storage is “HBM and other memory”. Networking is “Optical Circuit Switches, Datacenter Network & ultra-low latency ICI”.
TPUs are the AI chips Google designs. Google Cloud’s documentation describes TPU7x, the first release in its Ironwood family, as Google Cloud’s “seventh generation TPU” and lists 192 GiB of HBM capacity per chip.
What slide 10 of Alphabet’s Q3 2026 earnings slides lists inside a TPU system, redrawn: purpose-built hardware, components from a four-chip tray to a 144-rack pod, and software and support. Chip counts per rack and pod are our arithmetic, assuming one four-chip tray per machine. Sources: Alphabet Q3 2026 earnings slides, 28 October 2026; Google Cloud TPU7x documentation.
Layer Two: From Tray to Pod
The slide builds the system in three steps. A TPU tray is “Four proprietary chips on one liquid-cooled board”. A TPU rack holds “16 TPU machines with battery backup”. A TPU pod is “144 TPU racks with 3 optical switch racks”.
The slide does not state a chip count. If each machine holds one four-chip tray, a rack holds 64 chips and a pod 9,216, by our arithmetic. Google Cloud’s documentation lists 9,216 chips per pod for TPU7x and says each TPU7x virtual machine contains 4 chips.
At 192 GiB per chip, a 9,216-chip pod would carry about 1.77 million GiB of HBM, also our arithmetic.
Layer Three: Software and Support
The top layer is software and service. The slide lists frameworks with “Native support for JAX, PyTorch & XLA compiler”, an inference engine with “High-throughput vLLM, SGLang & model serving engines”, and “Post-sales product support and warranty”.
Thomas Kurian, chief executive of Google Cloud, said on the call that Google “recently worked closely with Meta to deliver Torch TPU which allows users to train, serve, and deploy PyTorch models on TPUs using the tools they already rely on.”
He described a system as “a combination of hardware, system software, compilers, the whole thing, assembled as a system, including optical networking and other things.”
Where the Systems Go
Kurian said: “In the past, you could use TPUs in our cloud.” For customers who want them on site, he said, “we’re making TPU systems available in their data centers.” He added: “We’re also broadening availability to other providers.”
He cited a partnership with Oracle and Crux AI, which he described as a joint venture, to offer customers TPU systems. He said “nearly 100 customers use TPUs today”. Tonight’s TPU customers piece lists the names he gave, and an earlier piece covered the first on-site sales.
Kurian said placing systems in third-party facilities and customer data centers allows Google “to broaden our addressable markets”, and that “it also offsets the need for us to prebuild facilities to host these TPU systems.”
How a Sale Turns Into Cash
Ashkenazi described the order of events. “When we sign an agreement with a customer to sell TPU systems, once we sign that agreement, you’ll first see that showing up in our backlog. And then we start building inventory.”
She said: “We build the inventory ahead of getting the revenue stream in.” Then, “as we start delivering and billing customers for these TPU systems, then you see cash coming in.”
Kurian said: “When Anat mentioned that we procure components to build inventory, we only do that after we have signed contracts, so all of the inventory for TPU systems are associated with committed contracts.” Alphabet’s balance sheet shows inventory of $13.3 billion at 30 September, against $2.4 billion at the end of 2025, as tonight’s inventory piece sets out.
Ashkenazi said: “To date we’ve had some system sales recorded as revenue, but expect more of it to come in through 2027.” In her outlook she said Google expects “TPU system sales to continue to scale in Q4, followed by a meaningful revenue ramp in 2027.”
The Next Chips
Kurian said: “TPU 8T and 8I offer up to 2.7X better training performance and 80% better inference price performance, respectively, versus our previous generation of TPUs.”
As we read the slide and the call, a TPU system sale bundles hardware, the components that scale it from tray to pod, and software and support; Alphabet’s release says Google Cloud generates product revenues primarily from the sale of TPU systems.
Ashkenazi said on the call: “There is a significant added value that’s being provided to the customer that’s purchasing a TPU system.”
By our arithmetic, Alphabet’s inventory was about 5.5 times its end-2025 level at 30 September: $13.3 billion against $2.4 billion.
“So it’s not the traditional kind of capex as you’re thinking about this”
Three Implications
FOR CUSTOMERS Kurian said on the call that Google is making TPU systems available in customers’ data centers and through other providers, including Oracle.
FOR ALPHABET’S CASH FLOW Ashkenazi said a signed sale shows up first in backlog and then in inventory, and that cash comes in as Google delivers and bills.
FOR THE CHIP ITSELF Kurian said TPU 8T and 8I offer up to 2.7X better training performance and 80% better inference price performance, respectively, than the previous generation.
The framework argues: “It is a layered industrial stack — and at any given moment, the binding constraint sits at a different layer than it did six months ago.”
Alphabet’s slide describes a TPU system in layers too: chips, memory and networking, then trays, racks and pods, then software and support. The framework does not discuss TPU systems.
What Is Not Established
Alphabet gives no separate revenue figure for TPU systems and no price per system. Ashkenazi said: “We don’t provide specific financials at the product level”, and Kurian said: “We don’t break out margins for individual products or infrastructure components.”
The slide does not say how many trays sit in one machine, so the per-rack and per-pod chip counts are our arithmetic on that assumption. It does not say which customers bought whole pods.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Alphabet’s slide describes a TPU system as three layers: purpose-built hardware, components that scale from a four-chip tray to a 144-rack pod, and software and support. Its executives said on the call that a signed sale shows up first in backlog, then as inventory, and later as revenue and cash, with more of that revenue expected through 2027.
Business Engineer Premium
Premium: Alphabet & Microsoft on the Map of AI
Cash conversion, the lengthening backlog, the arms-dealer turn at Layer 3: the full Business Engineer read of both prints.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The system description comes from slide 10 of Alphabet’s third-quarter 2026 earnings slides; call quotations come from a caption transcript of the 28 October 2026 webcast; inventory and segment wording come from Alphabet’s results release; chip specifications come from Google Cloud’s TPU7x documentation. Chip and memory totals are our arithmetic; linked earlier pieces carry their own sources.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
