Google Cloud Turned Into a Chip Channel: 20% Margins on Its Own Silicon, 80% at the Memory Suppliers

Google Cloud did not add a reseller this week; it converted a captive product into a sold one. Alphabet's Q2 2026 10-Q states that in the second quarter of 2026 it began recognizing revenue from the sale of TPU systems, and the same filing adds risk-factor language on supplying TPU systems to meet customers' AI compute demand. The Q2 earnings release is blunter still, describing Cloud product revenue as coming primarily from those system sales. Hardware sales are displacing rented compute inside the segment.

Run the mix and the drag appears. Wells Fargo expects roughly $9 billion of TPU sales in the third quarter against a segment that reported $24.8 billion in the second; at a 20% margin on the chips and last quarter's 35.5% on everything else, the blended number lands near 30%. That dilutes the segment margin by roughly 560 basis points from Q2. Treat it as a scenario rather than a forecast: the sales figures are billings estimates, and Alphabet's CFO said in July that only a small portion of TPU revenue from existing agreements is recognized this year, with the vast majority landing in 2027.

Broadcom is the clearest beneficiary of the shift. Fiscal Q3 2026 revenue of $29.6 billion was up from $16.0 billion a year earlier, semiconductor solutions revenue jumped to $20.8 billion from $9.2 billion, and gross margin held at 69%. Management described unprecedented demand for custom accelerators and AI networking, and the June 2026 launch of its AI XPU platform came with an initial $35 billion tranche aimed at enabling more than one gigawatt of customer compute. The same filing shows the concentration this creates: one semiconductor solutions customer, described as a distributor, was 50% of net revenue, and the top five end customers were about 55%. Growth at that concentration is growth in one relationship.

Marvell shows how far the channel reaches. Its Aug 19, 2026 filing discloses a commercial agreement with Alphabet signed July 29, 2026 for custom semiconductor products, and a warrant for up to 58,970,907 shares at $206.58, about $12.2 billion, that vests in 240 equal tranches, one for every $500 million of custom products revenue. The scope runs past the accelerator into AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute. Alphabet is now buying the components around its own chip, and Marvell's 16.7% operating margin shows that this layer does not earn design-layer pricing power.

Equity is not what funds this buildout; borrowed money is. On Oct 26, 2026, Bank of America, Citigroup, Morgan Stanley and Blackstone lined up $63 billion to buy Broadcom's AI accelerators and networking equipment and lease them to Anthropic, with deliveries starting in 2027. About $43.5 billion is first-lien debt arranged by the three banks, Blackstone holds the committed second-lien tranche, and Broadcom provides residual value support to the lenders. The structure is the point: the chip vendor guarantees the residual value while the banks own the asset and the customer rents the capacity.

Alphabet now runs the same trade a layer down. Its June 30, 2026 balance sheet carried $811.0 billion of purchase commitments and other contractual obligations, $200.7 billion of it current, plus an agreement to provide an estimated $24.1 billion of future backstops to support data-center and energy supply infrastructure. First-half capital expenditure was $80.6 billion against $39.6 billion a year earlier, and full-year guidance was raised to $195 billion to $205 billion. In the second quarter the company generated $39.1 billion of operating cash flow against $44.9 billion of capital spending, a negative $5.9 billion of free cash flow in the very quarter Cloud revenue grew 82%. It funded the gap with $49.6 billion of net equity proceeds and $20.3 billion of senior notes.

The buyer list behind the TPU channel now reads like a competitor's customer list. Anthropic committed in October 2025 to up to one million TPUs in a deal worth tens of billions and expanded it in April 2026 to multiple gigawatts of capacity with Broadcom and Alphabet. On Oct 28, CNBC confirmed that OpenAI plans to buy TPUs from Alphabet. And on Oracle's cloud, the menu is deliberately wide: NVIDIA Vera Rubin NVL72 systems and Vera CPUs already generally available with multi-gigawatt plans, TPUs from 2027, Cerebras wafer-scale inference for 2027, and OpenAI's own Jalapeño accelerators built with Broadcom. Choice is the product, and the buyer holds the negotiating position.

For scale, Amazon's AWS generated $42 billion of cloud revenue in the second quarter of 2026 and Microsoft's Azure $29 billion, against $25 billion at Google Cloud, which is growing 82% on hardware sales that rivals can now also rent. That is the strategic trap: the channel makes Alphabet's silicon reachable by customers it does not own, while making its cloud easier to compare against the alternatives. Alphabet sells a TPU to a customer who could have rented an NVIDIA GPU from the same rival cloud instead.

Near term, the test arrives immediately. Alphabet reports third-quarter results after the close on Oct 28, 2026, with Wall Street expecting Google Cloud growth of about 90% year over year. The number that matters is not the growth rate, it is the segment margin that has to carry it.

Over one to three years, the channel either becomes a second business for Alphabet or a second business for Oracle. The swing factors are all dated.

The answer to who keeps the margin is not a company; it is a balance sheet. The design layer keeps the engineering profit, with Broadcom at a 54.3% operating margin and NVIDIA at 66.2%. The memory layer keeps the scarcity rent, with SK hynix near 80% and Micron at 80.7%. The channel operator in the middle earns roughly 35% on rented compute, whether that is Google Cloud or Oracle, and about 20% on the silicon it manufactures. The reason to care is that the two highest-margin layers are also the two with the least substitutable supply. Alphabet's move converts a captive cost advantage into a market share gamble: it wins workloads from NVIDIA's rack-scale lock-in, and it gives away the margin that made those workloads worth having.

The practical test for the next four quarters is narrow and observable: does Google Cloud hold a mid-30s operating margin while a third of its revenue becomes hardware it manufactures, and does Oracle sign contracts for 2027 TPU capacity before Alphabet's own backlog converts? If the margin holds, the channel was a masterstroke of self-financing share capture. If it slides toward 30%, then Alphabet has simply become the cheapest way to buy a TPU, and the profit has already left the building for the designers, the memory makers and the lenders.