Alphabet’s Server Spend Faces a 15% Price Rise

Alphabet is putting about 60 cents of every technical-infrastructure dollar into servers this year. In July it lifted 2026 capital spending to $195 billion to $205 billion. Contract builders have now told large buyers, including Google, that Nvidia-based AI servers will cost more than 15% more on units shipped in early 2027.

That notice does not rewrite this year’s plan. It lands on 2027 racks, on a company whose server line is already the fat part of the budget, and on a hardware bill that Google’s own supply chief this week pinned on memory, not on the accelerator logo.

Sixty Cents of Every Infrastructure Dollar Goes to Servers

Chief financial officer Anat Ashkenazi walked through the mix on the July 22 earnings call. Second-quarter capital spending was $44.9 billion, with the vast majority in technical infrastructure for AI. Of that infrastructure slice, she said approximately 60% of technical infrastructure went to servers and 40% went to data centers and networking gear.

THE CAPEX RAMP

The 60/40 split is not a one-quarter quirk. Ashkenazi used the same words after the first quarter, and the year-end 2025 call used them too. Servers are the largest check inside a budget that is still climbing.

Approximately 60% of our investment in technical infrastructure this quarter was in servers, and 40% was in data centers and networking equipment.

Anat Ashkenazi, SVP and CFO, Alphabet Q2 2026 earnings call

The July raise added $15 billion to both ends of the 2026 range. Ashkenazi said the extra was mainly faster delivery of capacity, not a new price list. She also repeated that 2027 spending will “increase significantly,” with details later. The company is still, in her words, in a supply-constrained environment, with demand from Cloud customers and from Google’s own products running ahead of what it can install.

15% More for Racks Shipping Next Year

In late August, firms that build AI servers under contract told operators that prices are going up more than 15% in many cases. The notice covers systems built around Nvidia’s Vera Rubin and Grace Blackwell chips. Microsoft, Google, and Oracle sit on the customer list. The size of each increase depends on chip generation and on how much memory the configuration carries, and it applies to machines that ship in early 2027.

Nvidia has not posted a public price sheet for the change. The path is the usual one in this market: the chip designer sells into system builders, those builders assemble racks, and the racks go to the clouds. A percentage that large on machines that already cost several million dollars apiece is a lot of extra cash per hall, even before power, cooling, and the building.

WHAT THE PRICE NOTICE COVERS

• The systems: Racks built around Vera Rubin and Grace Blackwell chips.
• The timing: Units shipped in early 2027, not the servers Alphabet is installing this quarter.
• The size: More than 15% in many cases, not a single published list price.
• The lever: Chip generation and memory configuration, which is why denser HBM setups see a steeper bill.
• The buyers named: Contract builders supplying Microsoft, Google, and Oracle.

CEO Sundar Pichai already treats Rubin as part of Google Cloud’s shelf. On the same July call he said Cloud offers accelerators from Google and from Nvidia, “including the new NVIDIA Vera Rubin platform, and TPU 8t and 8i.” Google is not a bystander on the Nvidia line. It is a listed customer and a reseller of the same generation that is being repriced.

More Than 75% of the Server Bill Is Memory

On Tuesday at SEMICON Taiwan, Nikhil Cherian, senior director of supply chain infrastructure at Google Cloud, put a number under that mix. He told the memory forum that AI infrastructure has shifted from being limited by compute to being limited by memory, and that high-performance memory now accounts for more than 75% of an AI server’s hardware bill of materials. The conference program for the same talk framed memory as the dominant cost driver in that hardware.

That is the piece the 15% headline skips. A price notice on Nvidia systems reads like a GPU story. Cherian’s figure is a memory story. HBM and advanced DRAM sit beside the accelerator and feed it data. When those stacks take more than three-quarters of the parts cost, a 10% move in memory is already most of a 7% move in the server, before anyone reprices the chip.

He also described the bind in operational terms. Expensive accelerators idle when they wait on data, which wastes halls that Alphabet is filling at $44.9 billion a quarter. Google’s answer, as he laid it out, is split silicon and tighter software: TPU 8i for low-latency inference with 288 GB of HBM, TPU 8t for giant training pods, plus a quantization method that shrinks the working memory a model keeps in cache. He still said the software cut is not enough for the scale Google is planning.

The second-order point is blunt. Custom chips change who designs the accelerator. They do not create a private supply of HBM. A TPU board with 288 GB of high-bandwidth memory is still in the same three-vendor memory line that Nvidia is trying to lock up.

Does Alphabet’s TPU Mix Escape the Increase?

Only in part. Industry figures issued in February put TPUs at nearly 78% of AI servers shipped to Google this year, the only major cloud whose custom chips outnumber GPUs. That mix is real insulation against a pure Nvidia list-price shock. It is weak insulation against a memory shock, which is what Cherian described and what the 2027 system notices are tied to.

HOW THE BIG CLOUDS SPLIT AI SERVERS

Pichai was clear about the pecking order for scarce TPUs. Frontier model work comes first. Serving Search, Gemini, and enterprise products comes next. Customers who want TPU iron in their own buildings get what is left, including a project with Blackstone. Google also started delivering TPU systems to customer data centers in the second quarter, the first time those sales hit Cloud revenue. Ashkenazi said only a small slice of that contracted TPU revenue lands in 2026, with the vast majority in 2027.

So Alphabet is on both sides of the same constraint. It still buys Nvidia racks, including Rubin, for Cloud customers who want them. It builds and now sells TPU systems that need the same class of memory. And it is renting third-party capacity this quarter as a bridge, which Ashkenazi warned will put modest pressure on Cloud margins while internal halls catch up.

$279 Billion in Memory Nvidia Already Locked

Nvidia’s own accounts show why the 2027 sticker can move before the racks do. In commentary on the quarter ended July 26, the company said it raised supply commitments of $279 billion, up from $119 billion a quarter earlier, “primarily related to the procurement of memory.” Revenue in that quarter was $96.2 billion, up 106%, with data-center sales of $89.0 billion, up 117%. Inventory rose to $31.6 billion from $25.8 billion as it staged Vera Rubin.

Gross margin was 75.0%. Nvidia then guided the next quarter to 74.0%. The company is paying up for memory, booking the cost, and pushing system prices out to 2027 shipments. That sequence matches the notices contract builders sent in August.

Memory Nvidia has contracted through the next few years is memory someone else cannot buy. Google’s TPU line, Amazon’s Trainium boards, and Microsoft’s Maia parts all draw from the same short list of HBM suppliers. A pre-buy that large does not have to target Google to raise Google’s bill. It tightens the pool Cherian is already calling the limit on AI servers.

The public reason for the 15% is soaring memory costs. The practical result is a higher floor under racks that were already among the most expensive machines these companies buy. Waiting on another merchant GPU, or designing a new custom chip, does not mint extra HBM wafers this winter.

Paying for Servers With an 82% Cloud Jump

The demand side of the ledger is not theoretical. Alphabet’s second-quarter revenue was $119.8 billion, up 24%, or 23% in constant currency. Operating income rose 30%, and the operating margin reached 34%. Google Cloud was the spike: revenue of $24.8 billion, up 82%, with GCP growing faster than Cloud as a whole. Cloud operating income was $8.8 billion, more than tripling, and the Cloud margin widened to 35.6% from 20.7% a year earlier.

CLOUD AND CASH IN Q2

• Cloud backlog: $514 billion, with a bit more than half expected as revenue over the next 24 months.
• Cash from operations: $39.1 billion in the quarter and $185.7 billion over 12 months.
• Free cash flow: Negative $5.9 billion in the quarter, after $44.9 billion of capital spending; $53.3 billion over 12 months.
• Balance sheet: $242.5 billion in cash and marketable securities, with long-term debt of $98.2 billion.

Services still pay most of the bills. That segment took in $95 billion, up 15%, with Search & Other up 17% and YouTube ads up 13%. Other cost of revenues rose 22% to $29.8 billion, driven by depreciation, inventory costs from TPU systems built for customers, and YouTube content. Operating expenses rose 27% to $33.1 billion. The AI build is already in the P&L, not only in the cash-flow statement.

Ashkenazi said free cash flow will stay under pressure while the company keeps funding technical infrastructure. In June, Alphabet raised about $85 billion of equity to help pay for the same build, on top of operating cash and debt. The Cloud jump is the commercial case for that spend. Negative free cash flow in a quarter when Cloud grew 82% is the cash case that the 2027 price hike will aggravate, not create.

The Depreciation Stack Already in Motion

The 15% on 2027 shipments is a future cash item. Depreciation is the present one. Ashkenazi told investors the rise in technical infrastructure will keep pressing the income statement through higher depreciation and through running costs such as energy. That path was already visible last year. Depreciation rose from $15.3 billion in 2024 to $21.1 billion in 2025, nearly $6 billion, or 38%, and the company said the 2026 growth rate would accelerate.

In June, Pichai told investors 2022 capital spending was about $31 billion and that 2026, then guided at $180 billion to $190 billion, would be six times larger than 2022 spending and double the prior year. July’s raise took the live range to $195 billion to $205 billion. The overwhelming majority, he said, is technical infrastructure. Ten years after Google’s first commercial TPU, the company still lists Nvidia GPUs as a core part of the accelerator set. The hedge is real. It is not a wall.

Servers bought this year start wearing down on the income statement long before a 2027 Nvidia rack shows up at a loading dock. A higher 2027 unit price then adds a second wave of cash out the door, and a second wave of depreciation behind it. Cloud can fund a lot of that if 82% growth and a $514 billion backlog hold. The memory share Cherian put above 75% is why the unit price is moving even for a buyer that already designs most of its own AI servers.

Alphabet’s 2026 guide assumed faster delivery, not a new memory tariff. The tariff is now on the calendar for early next year, and it hits the line that already takes 60 cents of every infrastructure dollar. That is the bill, whether the board says TPU or Rubin.

Disclaimer: This article is news reporting and analysis of public company remarks, filings, and industry figures. It is for information only and is not investment advice, a recommendation to buy or sell Alphabet, Nvidia, or any other security, or a forecast of 2027 capital spending or Cloud margins. Readers who may act on company spending, chip pricing, or cloud costs should consult a licensed financial adviser who can review their own holdings and time horizon. Figures and statuses reflect the cited company comments and documents as of September 2, 2026, and guidance, prices, and product mix can change.