Google and Amazon's own AI chips put pressure on Nvidia as its founder steps back
Research firm Counterpoint expects custom cloud chips to outship Nvidia-style GPUs next year, just as Nvidia's long-time chief leaves the CEO role.
The short version
- Counterpoint predicts custom AI chips (ASICs) will ship in larger volumes than GPUs next year.
- Google has launched four TPU systems within a year; Amazon's AWS is building its Trainium chips.
- Nvidia keeps an edge in top-end computing, its CUDA software and system integration.
- MediaTek says it has won a second cloud customer for custom chips, with revenue coming later.
- Acting CEO Kress faces investors for the first time at Nvidia's next earnings call.
Why it matters
The chips behind AI models shape what running them costs. More serious suppliers usually means more bargaining power for big buyers over time, though that is a market pattern, not a promised price cut. For anyone buying AI services, it is another reason not to lock everything to one vendor.
A two-front squeeze on Nvidia
As Min Bao reported, Nvidia's co-founder Jensen Huang has just stepped down as CEO, and Google promptly pushed a new generation of AI chips into the market. Research firm Counterpoint expects cloud giants such as Google and Amazon's AWS to keep expanding their in-house chips.
Its headline forecast: next year, shipments of custom AI chips for servers will exceed shipments of GPUs. Counterpoint's Brady Wang made the prediction, and it counts chips shipped, not money earned or computing power.
Custom chips versus the general workhorse
A GPU is a flexible processor, originally built for graphics, that became the standard engine for AI. An ASIC is a chip designed for one narrow kind of work. Google's version is called the TPU.
Google announced two more TPU systems this month, adding to two from April, so four in about a year. It says the chips run its own Gemini models and are also rented to businesses through Google Cloud, or even placed in customers' own data centers. AWS is developing its Trainium line for training and running models.
The appeal for cloud companies is cost. They spend huge sums on data centers, and their own chips could lower the price of certain workloads and reduce dependence on Nvidia.
Nvidia is not out of the race
Counterpoint stresses that Nvidia still leads in top-end computing, in its CUDA software ecosystem (the tools developers build on) and in whole-system integration. The change is that big buyers have more choices.
The firm also sees cloud customers splitting in two. Some will move faster into custom chips. Others may slow spending because of power limits, data-center construction delays and pressure to show returns.
Taiwan's suppliers and a leadership test
Both camps need TSMC's advanced manufacturing and packaging, so it benefits either way. MediaTek said it has landed a second cloud custom-chip customer, with revenue arriving later. That puts it in competition with Broadcom. But more custom-chip orders do not automatically help every supplier, since each project has to be won on design.
Nvidia's roadmap from Blackwell to Vera Rubin is set, and Counterpoint thinks that, with the existing teams and supply ties, it can carry the company for a good while. The harder question is management. Dozens of senior executives reported directly to Huang, and the board is now searching for a new CEO.
Acting chief Kress, long the finance head, will face investors for the first time at Nvidia's next earnings call. Reuters reported she had previously raised objections to some large AI infrastructure financing commitments.
What to watch
- What Kress tells investors about demand and product timing at Nvidia's next earnings call.
- Whether the board's CEO search brings in an outsider or an insider, and whether key Nvidia staff stay.
- Whether cloud customers really split between speeding up custom chips and slowing AI spending.
Try it
Check whether your AI workflow depends on one vendor's special features. If it does, keep your prompts and data in portable formats so switching providers later is easy.
Sources and links
- ReportingMin Bao: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Min Bao · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
