Google Starts Selling TPU Chips, Turning Alphabet Into a Semiconductor Vendor

Alphabet has crossed a line it spent nearly a decade approaching. The company's custom Tensor Processing Units, built for years to run Google's own workloads inside Google's own facilities, are now being shipped to customer data centers. That single logistical change turns a search and advertising giant into something it has never formally been: a chip vendor selling silicon to other people.

The disclosure came during Alphabet's second-quarter 2026 earnings call on July 21, and it arrived alongside a set of numbers that explain both why the company is doing this and why the rollout will be slower than demand warrants.

Alphabet Books Its First TPU Hardware Revenue

Chief Financial Officer Anat Ashkenazi told analysts that TPU system deliveries to external data centers began during the second quarter, and that Alphabet started recognizing revenue from those sales in the same period. It is the first time hardware sales have shown up in the company's results this way.

The contribution was real but modest. Google Cloud revenue climbed 82% year over year to $24.8 billion, with TPU system sales forming part of that total. Ashkenazi was careful to set expectations on timing: only a relatively small share of revenue from existing TPU sales agreements will land in 2026, with the pace picking up toward the end of the year and the bulk of it recognized in 2027.

What the Backlog Signals

The more telling figure sits in Google Cloud's backlog, which absorbed the TPU hardware agreements and grew by $50 billion during the quarter to reach $514 billion. Backlog is contracted work not yet delivered, and a half-trillion-dollar figure suggests the constraint on TPU sales is not customer interest. It is Alphabet's ability to physically produce and ship the systems.

That distinction matters for anyone reading the revenue number and finding it small. The commitments have been signed. The chips have not all been built.

AGI Development Gets First Claim on TPU Supply

Asked directly how a supply-limited chip fleet gets divided, CEO Sundar Pichai did not soften the answer. Alphabet's top allocation priority, he said, is securing whatever compute the company needs to stay competitive at the frontier of artificial general intelligence development. Core product areas come next — he named Search, YouTube, and Cloud among them.

Read that order again, because it puts external customers at the back of the line. Anyone buying TPU systems from Google is buying capacity that Google has first decided it does not need for its own model training. In a market where every accelerator is spoken for, that is a meaningful term of trade, and it is one Nvidia customers do not face in the same form.

Ashkenazi outlined how the company plans to cover the shortfall in the near term. Google will lean more heavily on third-party capacity during the third quarter, treating it as a bridge while it continues building out its own infrastructure.

The TSMC Packaging Bottleneck Behind the Shortage

The supply problem traces to a specific point in the manufacturing chain. Taiwan Semiconductor Manufacturing Company fabricates Google's TPUs, and the limiting factor is not wafer production but advanced packaging — specifically CoWoS capacity, the step that bonds logic dies and high-bandwidth memory onto a single substrate.

That bottleneck has already forced a downward revision. Production targets for 2026 were cut from roughly 4 million units to around 3 million. A million missing chips is not a rounding error when each one carries a backlog commitment behind it.

The 2027 Ramp

Analysts expect output to recover sharply as TSMC brings more packaging capacity online, with projections in the range of 5 million to 6 million units in 2027. The timing lines up with Ashkenazi's revenue guidance, which is not a coincidence — the year Alphabet expects to recognize most of its TPU hardware revenue is the year it expects to be able to build the hardware in volume.

Alphabet is also working to widen its supply base beyond a single foundry partner. The company is developing a chip known as Icefish alongside MediaTek, with mass production targeted for as early as 2028, part of a broader effort to reduce dependence on TSMC's most advanced nodes as demand from Nvidia, AMD, Apple, and others keeps them running near capacity.

What Direct TPU Sales Mean for the Nvidia Rivalry

Until now, Google competed with Nvidia indirectly. Customers could rent TPU capacity through Google Cloud, but if they wanted accelerators inside their own facilities, Nvidia was the answer. Selling TPU systems into customer data centers removes that separation and puts the two companies in the same market, competing for the same rack space.

The competitive picture is more complicated than a straight head-to-head, though. Google is entering as a supplier that has publicly committed its best silicon to its own frontier research first, at a moment when it cannot build enough units to satisfy existing contracts. That is a difficult position from which to win share on availability.

The Cost of Building the Fleet

None of this is cheap, and the second-quarter results show the strain plainly.

Alphabet raised full-year 2026 capital expenditure guidance to a range of $195 billion to $205 billion, up from the $180 billion to $190 billion it projected only a quarter earlier. The company spent $80.6 billion on capital expenditures in the first half of 2026 alone, against $39.6 billion in the same period a year prior.

Funding that pace has pushed Alphabet into the capital markets at unusual scale. During the quarter, the company raised $49.6 billion through new stock and mandatory convertible preferred shares earmarked for AI infrastructure, plus $20.3 billion in senior unsecured notes.

The clearest signal of what all this costs: Alphabet's free cash flow turned negative for the first time, landing at minus $5.9 billion for the quarter. A company that has generated cash almost effortlessly for two decades is now spending faster than it earns in order to build the compute base it believes the next phase depends on.