Nvidia has warned major customers that Grace Blackwell and Vera Rubin server costs could rise more than 15% in early 2027 as memory prices climb. The increase hands Alphabet's TPU business a stronger sales pitch and adds momentum to Marvell's custom-chip pipeline, positioning both as beneficiaries regardless of which chip architecture wins out.
Some of Nvidia's largest customers have been warned that AI servers built around Grace Blackwell and Vera Rubin chips could cost more than 15% extra in early 2027 as memory prices climb. Alphabet is among the data center operators exposed to those increases, putting it in an unusual position: it buys Nvidia hardware for customers who want it while also selling a competing stack built around its own tensor processing units.
Alphabet leans on its TPU stack
Alphabet has committed $195 billion to $205 billion in capital spending this year on AI infrastructure, and a new partnership with Marvell could generate up to $120 billion in custom-chip sales through fiscal 2033. Google Cloud revenue grew 82% in the June quarter to $24.8 billion, and its cloud backlog reached $514 billion, with Alphabet expecting to recognize just over 50% of that backlog as revenue over the next 24 months.
Sundar Pichai framed the offering as accelerator-agnostic. According to 24/7 Wall St.: "the industry's broadest range of accelerators from Google and NVIDIA". That posture matters because Alphabet can absorb Nvidia price increases and monetize its own alternative at the same time.
The buildout has still come at a cost. Q2 free cash flow was negative $5.9 billion, long-term debt rose to $98.2 billion, and Alphabet raised roughly $70 billion in combined equity and debt to fund it. Yet investors have not punished the spending: GOOGL is up 69.37% over the past year and closed at $348.06 on August 24.
Marvell wins either way
Marvell reported Q1 fiscal 2027 revenue of $2.42 billion, up 27.57% year over year, with data center revenue accounting for 76% of the total. Q2 guidance implied roughly 35% annual growth. CEO Matt Murphy said Marvell has custom engagements at every major U.S. hyperscaler, and the custom chip business alone has a path to over $10 billion in revenue by fiscal 2029, based on designs already won rather than a speculative pipeline.
The company has also expanded its relationship with Nvidia around silicon photonics and NVLink Fusion, letting it build custom chips and networking semiconductors that interface with Nvidia infrastructure. The same optics and interconnect products sell into both TPU and Blackwell server racks, so Marvell's business does not depend on which architecture customers ultimately choose.
Who benefits most
A reported 15% surcharge on Nvidia servers probably is not enough on its own to trigger a mass migration, but it changes negotiating leverage. Enterprises renewing multi-year cloud contracts now have a reason to price out TPU capacity alongside GPU capacity. Alphabet does not need outside customers to justify its buildout — its own search, YouTube, and Gemini training workloads can absorb much of the capacity even if external TPU demand grows more slowly than the backlog implies.
Marvell supplies the optics, interconnect, and often the custom silicon itself regardless of whether hyperscalers lean on Nvidia GPUs or their own custom chips, which is why its stock has already outperformed.
Source: 24/7 Wall St.
Trading involves risk.
Risk Disclaimer
