Googles 175B AI Bill|Who Gets Paid Instead of Nvidia

The Day AI Money Moved Somewhere New

Nvidia fell on Monday. Not because of earnings, not because of a macro shock, and not because demand for AI chips dried up. It fell because one of its biggest customers may be working to replace it.

Alphabet is in advanced talks with Marvell Technology to co-develop two new versions of custom AI chips for Google Cloud workloads, according to reports circulating Monday morning. Marvell surged nearly 4 percent. Alphabet shares were roughly flat. Nvidia dropped about 1 percent — a quiet move, but one that carries a clear message.

This happened on the same day that JPMorgan named Alphabet its top pick ahead of the Google Cloud Next event in Las Vegas, projecting that Google Cloud could represent 19 percent of total Alphabet revenue in 2026. The bank noted that Google Cloud's backlog surged 160 percent year-over-year to $240 billion in late 2025. The AI infrastructure buildout is not slowing. If anything, it is accelerating. So the question is not whether Alphabet is spending. The question is who it is paying.

Alphabet's 2026 capital expenditure guidance sits between $175 billion and $185 billion. That is not a rounding error. It is a number that could reshape an entire semiconductor supply chain — and on Monday, the market got its first real hint of what that reshaping might look like.

When the Customer Becomes the Chipmaker

The logic of hyperscaler chip development is straightforward once you see the pattern. Google's AI workloads — inference, memory-intensive processing, large-scale cloud serving — do not need a general-purpose GPU. They need silicon built for exactly those tasks, at a cost structure that Nvidia does not offer.

Broadcom has been the dominant name in this space. Meta recently formalized a chip partnership with Broadcom for its own AI infrastructure. Now Alphabet appears to be pulling Marvell into the same orbit, and the market immediately marked Nvidia down on the news.

That is the signal worth watching: not Marvell up 4 percent, but Nvidia down 1 percent on a day when AI spending headlines were overwhelmingly bullish.

There is a historical parallel here. When Apple announced in 2020 that it was abandoning Intel processors for its own M1 chips, Intel shares fell and the move was initially dismissed as limited to consumer devices. Within two years, the M-chip transition had reshaped Apple's margin profile and Intel had lost one of its most prestigious customers permanently. The transition did not happen overnight. It happened in stages. And each stage was announced with the same framing: "this is for specific workloads."

The Alphabet-Marvell talks are at an early stage. No deal has been confirmed. But if the pattern holds, what starts as targeted chip co-development for inference workloads tends to expand. The question is not whether Nvidia loses Google entirely — it is whether the percentage of Google's AI compute budget flowing to Nvidia begins to shrink, and at what pace.

There is a condition that complicates this logic. Nvidia's ecosystem — its software stack, CUDA libraries, developer tooling — creates switching costs that custom silicon cannot easily replicate overnight. Alphabet has built its own TPUs for years and still uses Nvidia GPUs for many workloads. In-house and third-party silicon have coexisted. The threat to Nvidia is not replacement. It is market share compression.

What the Next 90 Days Will Reveal

Two events in the near term will clarify how real this shift is. The first is Google Cloud Next in Las Vegas, where JPMorgan expects the company to focus on what it calls "Agentic Cloud" — AI agents running multi-step workflows at scale. If Alphabet uses that stage to announce custom silicon milestones or deeper infrastructure independence, the Marvell talks will look like the opening move in a much larger repositioning.

The second is Alphabet's upcoming earnings report. If management discusses capex allocation in terms of custom silicon versus third-party procurement, that commentary will be more important than any chip announcement. A shift in language — from "we use Nvidia GPUs" to "we are building for our specific workloads" — would be a structural signal, not a one-day stock story.

The current evidence leans in one direction. Google Cloud's $240 billion backlog creates enormous pressure to reduce per-unit compute costs. Custom chips, once developed, cost less per inference than off-the-shelf GPUs at scale. The financial logic for in-housing is strong, and it compounds over time as cloud revenue grows.

The scenario that breaks this logic is Nvidia's own roadmap. If Blackwell and the generations that follow it close the efficiency gap between general-purpose and custom silicon — or if Nvidia's software advantages prove harder to replicate than expected — the case for co-developing proprietary chips weakens. Enterprise customers buying into Google Cloud's agentic AI stack may also demand workload portability, which keeps third-party hardware in the picture longer than the bulls on custom silicon expect.

Nvidia's 1 percent decline on Monday was not a collapse. It was a question. The answer depends on how many of its largest customers are having the same conversation that Alphabet is reportedly having with Marvell right now.