What Does Google Do? The TPU Conversion Model

Google books Cloud backlog, spends ads cash to stand up TPU and GPU halls, and gets paid as customers use live capacity or take delivery of TPU systems — demand still outruns what is installed, so watts are the stuck step.

HYPERSCALERS · Data as at 22 Jul 2026 · Next update: after Alphabet Q3 2026 results (28 Oct 2026)
Affects: Hyperscalers · Neoclouds · US

It is a hyperscaler in the AI build chain: a large cloud operator whose capital spending pulls chips, packaging, halls, and power. The AI cash, and the stuck step, sit in Google Cloud. Alphabet is the filer; Google is the operating name on this layer. For the book, see remaining performance obligation, explained. For the chip, see what is a TPU. For a peer conversion model, see What Does Amazon Do?.

In short: Google books Cloud, funds halls from cash capex, and invoices as customers use live capacity — or, new this quarter, as TPU systems are delivered. In the quarter ended 30 June 2026, Google Cloud sales were $24.8bn (+82%), Cloud operating income was $8.8bn, and Cloud remaining performance obligations were $513.9bn. Anat Ashkenazi said demand still outpaces the add, 2026 capex is now $195–205bn, and the vast majority of existing TPU-system sales land in 2027. The model works if racks and watts light on a clock the book can convert — not if $514bn is this year’s cash.

Education only, not advice to buy or sell any security. Figures are from company filings and company releases dated as shown. Secondhand reports are labelled reported.

Numbers that matter

Sources: Alphabet Q2 2026 results (company), 22 Jul 2026; Form 10-Q for the quarter ended 30 June 2026; Q2 2026 earnings call the same day. The $514bn Cloud backlog on the call is the 10-Q’s $513.9bn, rounded. The $195–205bn capex range is spoken on the call, then restated as the company’s guide.

What does Google actually do?

Three steps: Book. Light. Bill.

1. It books Google Cloud. Alphabet sells Search, YouTube, subscriptions, devices — and, on this layer, Google Cloud. The 10-Q put remaining performance obligations (“revenue backlog”) at $519.5bn as at 30 June 2026, of which $513.9bn related to Google Cloud. That is contracted work not yet recognised as revenue. It is not cash in the bank. The same note says just over 50% should be recognised over the next 24 months, driven by duration, delivery, and use. In Q1 2026 Alphabet changed the print to include contracts of one year or less, so this book is not the same cut as Amazon’s long-dated AWS line. See remaining performance obligation, explained.

2. It lights halls and accelerators. That is the operating job that binds. Ashkenazi said on the 22 July 2026 call that capacity is up “quite significantly” over three years, “the demand still outpaces that investment,” and that Google is “working in a supply-constrained environment.” Q2 purchases of property and equipment were $44.9bn, “the vast majority” technical infrastructure. Full-year 2026 capex was raised to $195–205bn from $180–190bn to accelerate delivery, with another step-up expected in 2027. The 10-Q recorded $80.6bn of property-and-equipment purchases in the first half, and $707.0bn of remaining contractual commitments — mostly long-term supply agreements through 2030, plus energy take-or-pay terms running as far as 2054.

3. It bills live Cloud — and, new this quarter, TPU systems. Google Cloud sales grew 82% to $24.768bn. Operating income was $8.814bn, more than triple the year-earlier $2.826bn; the call put the Cloud margin at 35.6%. GCP grew faster than Cloud overall. The company also began to recognise revenue from TPU system sales delivered into customer data centres for the first time in Q2. Ashkenazi said Cloud still accelerated after excluding those sales, that Q2 was a small slice of the existing TPU agreements, and that the vast majority of those agreements is recognised in 2027. Pichai said most cloud serving already runs on TPUs, with external demand balanced across TPUs and GPUs. See TPU vs GPU and GPU as a service, explained.

How does Google make money?

It gets paid when ads, subscriptions, and — this cycle — delivered Google Cloud are recognised as revenue. Three things decide how fast the Cloud line can grow:

• Delivery — a booked megawatt earns nothing until the hall is live
• Utilisation of what is live — Ashkenazi said demand still outpaces the add, so extra seats are not spare
• Mix and conversion — remaining performance obligations become revenue on usage, and TPU-system sales become revenue on delivery

On 22 July 2026 Alphabet printed $119.796bn of revenue, up 24%. Operating income was $40.770bn (34% margin). Unlike CoreWeave, it is still profitable on a GAAP operating basis while it stands up AI halls. Net income of $112.193bn is the wrong Cloud clock: other income was $97.983bn, including a $99.031bn gain on equity securities that the release says lifted net income by $77.1bn. The 10-Q attributes the jump primarily to unrealised gains on SpaceX and a private company. SpaceX shares on the books were $80.0bn (short-term sale restrictions) plus $14.1bn (locked through Q3 2027). Paper marks do not light a hall.

Our arithmetic. Cloud was 20.7% of sales (24.768 ÷ 119.796 = 0.207), 21.6% of operating income (8.814 ÷ 40.770 = 0.216), and 35.6% of its own sales (8.814 ÷ 24.768 = 0.356), matching the call. Google Services was still $94.540bn of sales — 78.9% of the company (94.540 ÷ 119.796 = 0.789). Ads still pay for the build.

Cash tells a second story. Trailing-twelve-month operating cash flow was $185.675bn less $132.402bn of property-and-equipment purchases, so free cash flow was $53.273bn. Our check: 185.675 − 132.402 = 53.273. Q2 itself ran the other way: $39.069bn of operating cash flow less $44.924bn of capex is −$5.855bn, which the company rounded to −$5.9bn. H1 purchases were $80.598bn. A $200bn midpoint on the $195–205bn guide less $80.6bn leaves about $119.4bn implied for the second half (200 − 80.6 = 119.4). See cash capex, explained.

Who gets paid: Alphabet, for live Google Cloud and for the ads that fund the hall. Money sticks at energised, accepted capacity — and, on the TPU-system lane, at delivery into a customer’s hall. The remaining-performance-obligation print is a claim on future usage and future hardware handovers.

What limits it?

Delivery first. Power and sites second. TPU mix third.

1. Demand still outpaces available capacity. Adding $44.9bn of capex in a quarter did not create spare seats. Ashkenazi said the company will expand third-party capacity in Q3 as a “bridging strategy” while it builds more internal halls, and that this will put modest pressure on Cloud operating margins. Pichai’s first priority on TPUs is still Google’s own products; external Cloud customers get a mix of TPUs and GPUs. See speed-to-power and the US market door.

2. Remaining performance obligations are not a shipping schedule. Timing is driven by duration, delivery, and use. Our arithmetic: $513.9bn against a $99.1bn Cloud run rate ($24.768bn × 4) is about 5.2 years of current Cloud sales (513.9 ÷ 99.072 = 5.19). “Just over 50%” in 24 months is about $257bn (0.50 × 513.9 = 257), or $128bn a year versus a $99bn run rate. The gap is the conversion ramp, including TPU-system sales the call puts mostly in 2027. $514bn is not $514bn of 2026 Cloud.

3. Watts are the long clock under the same bind. Three dated power prints sit next to the Q2 book. On 6 Oct 2026 Google and Constellation announced a 20-year deal to add 890 MW of nuclear uprates on PJM (first uprate by 2028, all by end-2032) plus 2,700 MW of existing generation. The same day Black Hills said it will spend $1.8bn on 564 MW of gas to serve a planned Google Cheyenne hall, with service late 2027 and peak load in 2030. On 7 Oct 2026 Indiana’s IURC approved AES Indiana’s plan for Google’s 390 MW Monrovia hall. Those are power clocks, not this quarter’s Cloud invoice. See the Wires on Constellation, Black Hills, and Monrovia, and the PJM capacity-short US door.

4. Capex is the cash cost of the same bind. Our arithmetic: $205bn at the top of the guide less first-half cash capex of $80.6bn leaves about $124.4bn for the second half if the top of the range holds (205 − 80.6 = 124.4). Q2 free cash flow is already negative. Ads cash is the funder; watts are the gate.

The honest uncertainty is conversion. A long book, a short hall, and a 2027 TPU-system clock are different facts.

Who else does this?

Roles, not a league table.

What separates them is who can deliver accepted megawatts, who funds the hall, whether the chip is merchant GPU or custom silicon, and whether demand still exceeds what is live after the add.

What would change this view?

The read on the model weakens if:

• Cloud growth slows while remaining performance obligations keep rising
• Ashkenazi’s “demand still outpaces” language disappears in a dated print
• Calendar 2026 capex is cut inside the $195–205bn range
• TPU-system revenue stays a rounding error into 2027, contrary to the call
• A named power print — 890 MW of uprates, Cheyenne, or 390 MW Monrovia — slips
• Hyperscaler peers cut 2027 capex — the demand-dial risk on the explained page
• Next GPU attach is late because packaging or HBM is short — the stuck step would then sit in semiconductors

It stays intact while demand outpaces live capacity after large capex adds, Cloud keeps converting, TPU-system sales follow the 2027 clock, and the $514bn book remains a usage clock rather than this year’s cash.

Practical takeaways

• Google’s AI model on this layer is Book → Light → Bill
• Q2 2026: Cloud $24.8bn (+82%), operating income $8.8bn, remaining performance obligations $513.9bn, capex $44.9bn, demand still outpaces supply
• GAAP operating income $40.8bn. Do not use $112.2bn net income as the Cloud clock
• Q2 free cash flow −$5.9bn; TTM still $53.3bn; 2026 capex $195–205bn
• The stuck step is capacity and power. TPU-system sales started; most of those agreements are a 2027 clock
• Next hard read: Q3 2026 results on 28 Oct 2026

FAQ

What does Google do?
Google runs Search, YouTube, subscriptions, and Google Cloud. The scarce job on this layer is delivering Cloud capacity — and TPU systems — then converting remaining performance obligations into recognised cloud revenue.

What does Google do in simple terms?
It books long Cloud contracts, switches on data-centre megawatts and custom chips, and charges once the rack is running. Some customers now buy the TPU hardware itself.

Is Google profitable?
Yes on a GAAP operating basis: $40.770bn of operating income on $119.796bn of revenue in the quarter ended 30 June 2026 (22 Jul 2026). Cloud operating income was $8.814bn. Trailing-twelve-month free cash flow was $53.3bn; Q2 free cash flow was an outflow of $5.9bn. Net income of $112.193bn includes a $77.1bn equity-securities lift and is not the Cloud print.

How is Google different from a neocloud?
A neocloud sells AI compute as the business. Google sells a broad cloud and still earns from ads; Cloud is the line growing fastest on this layer and the step that is stuck. Google also designs TPUs and has started selling TPU systems. See neocloud vs hyperscaler and What Does CoreWeave Do?.

What is Google Cloud remaining performance obligation?
$513.9bn of contracted Google Cloud performance not yet recognised as revenue as at 30 June 2026, out of $519.5bn for Alphabet. The 10-Q says just over 50% should be recognised over the next 24 months. The call’s $514bn Cloud backlog is that number, rounded. It is a backlog clock, not this quarter’s cash.

Should I buy Alphabet shares?
We do not make buy or sell calls, and nothing here is a recommendation. What we can tell you is what the company does, how it gets paid, and which figures would change the picture. The decision is yours.

Keep reading on the map

• The layer: Hyperscalers — capex pull and the delivery bind
• The term: Hyperscaler, explained
• The mechanism: Remaining performance obligation, explained · What is a TPU · TPU vs GPU
• The split: Neocloud vs hyperscaler
• The peers: What Does Amazon Do? · What Does Microsoft Do? · What Does Oracle Do?
• The Wire: Google signs 890 MW of nuclear uprates · Black Hills to build $1.8bn of gas for Cheyenne · Indiana clears AES for a 390 MW Google hall
• Where it happens: the US — 6.8 GW short on PJM · speed-to-power

The members cut on the long clock is already live: Where does the stuck step move?. This page stays the free Stack read of the Google model.

Sean

Writes Second Order by TKN — a plain-English map of AI infrastructure. Who gets paid, where money sticks, which step is stuck. Not investment advice.

How we check this

Figures come from company filings and company releases, dated below. Capacity, TPU mix, and the capex range from the earnings call are the company’s spoken print. Power deals are from the named company or regulator documents, not from secondhand reports. Arithmetic is shown in full.

Last reviewed: 13 Oct 2026 · Next review: after Q3 2026 results (Alphabet has called the print for 28 Oct 2026). Editorial standards →

Sources

1. Alphabet, Q2 2026 results (Exhibit 99.1) (company), 22 Jul 2026 — $119.8bn revenue, $40.8bn operating income, Cloud $24.8bn (+82%) and $8.814bn operating income, $99.031bn equity-securities gain lifting net income $77.1bn, FCF table
2. Alphabet, Form 10-Q for the quarter ended 30 June 2026 — Cloud remaining performance obligations $513.9bn of $519.5bn; just over 50% in 24 months; H1 PPE $80.598bn; commitments $707.0bn; SpaceX $80.0bn + $14.1bn
3. Alphabet, Q2 2026 earnings call transcript (company), 22 Jul 2026 — $514bn Cloud backlog; TPU systems first delivered in Q2, vast majority of those agreements in 2027; demand still outpaces capacity; 2026 capex $195–205bn
4. Alphabet, Alphabet Announces Date of Third Quarter 2026 Financial Results Conference Call (company), 6 Oct 2026 — Q3 print 28 Oct 2026
5. Constellation Energy / Google, 890 MW of nuclear uprates on PJM (company), 6 Oct 2026; also Google’s blog
6. Black Hills Corp., Cheyenne agreements (Exhibit 99.1) (company), 6 Oct 2026 — $1.8bn, 564 MW nameplate, service late 2027
7. Indiana Utility Regulatory Commission, Cause No. 46394 (regulator), 7 Oct 2026; AES Indiana (company) — Monrovia hall