Google is already selling AI chips. OpenAI reportedly wants to buy them

Alphabet began booking TPU system sales in the second quarter, and its inventory has quadrupled to $10 billion. A report that OpenAI plans to buy chips from Google would move the cost of owning them from Google's cloud to the customer.

Google already sells its AI chip as a product, and an OpenAI purchase would hand the cost of owning it to OpenAI. The Information reported on Tuesday evening that OpenAI plans to buy Google's AI chips, citing a person familiar with the matter. OpenAI is one of Nvidia's largest customers. Alphabet reports its third quarter today.

The chip is the tensor processing unit, or TPU. Alphabet's 10-Q for the quarter ended 30 June says Google Cloud "generates product revenues primarily from the sale of TPU systems". The filing describes supply agreements with customers who "require or provide on-premises infrastructure".

Alphabet began recognising revenue from those agreements in the second quarter. The filing says the "significant majority" will be recognised in 2027. A deal with OpenAI would move chip ownership, depreciation and the risk of idle hardware from Google's cloud to OpenAI. The larger change is that Google becomes a merchant chip seller.

Inventory and cash costs

The balance sheet shows the hardware build-up. Inventory rose from $2.4 billion at the end of 2025 to $10.0 billion on 30 June. That is $7.6 billion more, a 4.1 times increase in six months.

The inventory note describes the stock as hardware for TPU systems "and devices", so the $10.0 billion is not all chips. It still shows a company stocking up to ship products.

The cash flow statement shows the same move. Inventory absorbed $7.7 billion of operating cash in the first six months, against $0.6 billion a year earlier. Google is paying for hardware before it books the sale.

Alphabet's revenue note says Google Cloud product sales generally consist of TPU systems. These comprise hardware, software, installation, support and extended warranty services. Hardware revenue is recognised generally when control passes to the customer. Some arrangements include options accounted for as rights of return.

Google's price list shows what committing is worth. For one Ironwood chip in Iowa, Google's list prices run from $12.00 an hour on demand to $8.40 on a one year commitment. That is 30% below on demand. A three year commitment costs $5.40, which is 55% below on demand.

At 100,000 chips for three years, that commitment would cost $14.2 billion at list. That is a rental price quoted for one region, and large customers negotiate. A buyer that takes the hardware instead moves depreciation and idle capacity from Google's books to its own.

Who carries the hardware risk

Renting keeps the chips on Google's books. Alphabet spent $44.9 billion on capital expenditure in the second quarter and $80.6 billion in six months. A sale turns part of that spending into inventory and product revenue recognised on delivery. The buyer then carries depreciation and utilisation, and pays off only if the chips run full for their working life.

Google has reason to want a second sales channel. Google Cloud revenue rose 82% to $24.8 billion in the second quarter. Operating income rose 3.1 times to $8.8 billion.

Cloud profitability is already high. Its operating margin reached 35.6%, from 20.7% a year earlier. A product line that adds revenue without a rental contract extends that growth.

Demand is contracted. Google Cloud's backlog is $513.9 billion, of $519.5 billion in total. Alphabet expects just over half of the total within 24 months.

The new channel has a cost. The filing warns that supplying TPU systems "may increase our costs and operational complexity". The chips also leave Google's own data centres.

Alphabet already backstops other companies' capacity. It carries $7.6 billion of lease guarantees and credit derivatives with a $43.8 billion maximum. Its purchase commitments total $811.0 billion, $200.7 billion of it short term.

Customers gain more chip suppliers

Oracle shows the same direction from the cloud side. Its announcement on Tuesday says TPUs will be available through Oracle Cloud Infrastructure in calendar 2027. It calls this the first time they are offered on a hyperscale cloud outside Google, ramping over a multi year period. That is the same year the filing says most TPU system revenue is recognised.

Oracle is adding Google TPUs, Cerebras and SambaNova while expanding its Nvidia supply, including Vera Rubin systems, according to Reuters. Buyers are spreading their chip sources.

Google's TPU page says one cluster can hold up to 1 million chips. Nvidia remains the scale player. Its Data Center revenue was $89.0 billion in the quarter to 26 July, up 117%. It guides to $108.0 billion of total revenue for the next quarter.

OpenAI is not leaving Nvidia. Nvidia is the exclusive AI compute provider at the PORTS-Pike campus in Ohio, where OpenAI is the customer for 8 IT GW. Nvidia also put $1.5 billion into SB Energy, the campus developer.

A TPU purchase would compete for OpenAI's capacity beyond that campus. Reuters, which relayed the report, adds that OpenAI began renting Google's chips in June 2025. It also says deals with customers including Anthropic have fuelled TPU demand.

OpenAI unveiled its first custom chip with Broadcom in June and planned to deploy it by the end of 2026. That makes Google a third source of supply. A purchase would move an existing relationship from rented capacity to owned hardware.

What to watch in Alphabet’s results

Two items in today’s results will test this analysis. The first is inventory. The balance sheet at 30 September will show whether the $10.0 billion held on 30 June has grown. Growth would mean more TPU systems are being built for delivery in 2027.

The second is management's wording on 2027 TPU revenue, the year the filing says the significant majority of supply agreement revenue arrives. An OpenAI purchase would first show up there, as a larger 2027 figure, not in Cloud margin.