NewsStocksJPMorgan Puts Nvidia's Share of Future AI Chip Deals at 39% as AMD and Rivals Gain Ground

JPMorgan Puts Nvidia's Share of Future AI Chip Deals at 39% as AMD and Rivals Gain Ground

Author: Hokanews·

Key Takeaways

  • •A JPMorgan estimate places Nvidia's share of future compute deals with frontier AI laboratories at 39%, down from more than 80% during the 2023-2025 period.
  • •Broadcom custom chips are projected to account for 24% of those future deals, with Google TPUs at 16%, AMD at 15% and Amazon Trainium at 6%.
  • •One frontier AI lab says moving its models to AMD chips now takes three days instead of an earlier estimate of six to 12 months, crediting its own AI for performing the engineering.
  • •Six of the seven largest AI chip buyers are developing their own hardware, with SpaceX cited as the only exception.
  • •JPMorgan's note, released two days after Jensen Huang resigned as CEO on November 12, also reported that Nvidia devoted 58% of its operating cash flow to equity investments and acquisitions over the last three quarters.
JPMorgan Puts Nvidia's Share of Future AI Chip Deals at 39% as AMD and Rivals Gain Ground

Nvidia's grip on the computing infrastructure used by leading artificial intelligence developers is facing growing competition, with a JPMorgan estimate placing the chipmaker's share of future compute deals among frontier AI laboratories at 39%. That compares with a market share of more than 80% between 2023 and 2025, according to figures shared by Coin Bureau on X. The figures point to a more fragmented landscape for AI computing than the one that prevailed during the earlier period. The composition of that landscape matters because the underlying hardware shapes the cost, availability and software tooling surrounding large-scale AI training.

The remaining share of future deals is distributed across several competing chip platforms. Broadcom custom chips account for 24%, Google Tensor Processing Units (TPUs) for 16%, AMD for 15% and Amazon Trainium for 6%, a breakdown that highlights the range of alternatives being pursued by major AI developers as they diversify their computing infrastructure. The platforms differ in origin: Broadcom's chips are custom accelerators built to individual customers' specifications, Google's TPUs and Amazon's Trainium are developed in-house by the operators that deploy them, and AMD's data-center offering centers on its Instinct line of AI accelerators. Because the figures are estimates of deals yet to be made, they describe a projected procurement mix rather than recorded transactions.

AMD Chip Transition Reportedly Takes Just Three Days

One of the developments cited in the report concerns how quickly an AI laboratory can move its models onto hardware from another supplier. According to the information shared by Coin Bureau, one frontier AI lab said switching its models to AMD chips now takes three days, compared with an earlier estimate of six to 12 months.

The lab attributed the shorter transition period to its own AI performing the engineering work required for the switch. The reported change illustrates how the use of AI in engineering processes can reduce the time involved in adapting models to different chip architectures. Such transitions have historically demanded substantial software work, since models must be adapted to each vendor's instruction set and supporting tools before they run efficiently on new hardware.

The development is relevant to Nvidia's competitive position because the ability to move workloads between hardware providers can give AI laboratories more flexibility when selecting computing infrastructure. The source, however, does not identify the laboratory or specify which AMD chips are involved.

Major AI Chip Buyers Are Developing Their Own Hardware

The report also points to efforts by leading AI companies to develop their own chip technologies. Of the seven largest AI chip buyers, "only SpaceX is not trying to build their own," according to the statement cited by Coin Bureau.

That trend adds another dimension to competition in the AI infrastructure market. Alongside purchases from Nvidia, AMD and other suppliers, some of the largest buyers are pursuing in-house hardware development rather than relying entirely on external chip providers for their computing needs. The pattern is already visible in JPMorgan's breakdown, where Google's TPUs and Amazon's Trainium — silicon developed by the companies that deploy them — sit alongside chips purchased from merchant suppliers such as Nvidia and AMD.

Nvidia's Investment and Acquisition Spending

JPMorgan's note also highlighted Nvidia's capital allocation. The company spent 58% of its operating cash flow on equity investments and acquisitions over the last three quarters, according to the figures reported by Coin Bureau.

The timing of the note was also notable: it was published two days after Jensen Huang resigned as CEO on November 12, as stated in the original Coin Bureau post. The report did not provide further details about the circumstances of the resignation.

Taken together, the figures point to a changing competitive landscape for AI computing. JPMorgan's estimate places Nvidia at 39% of future frontier AI lab compute deals, while Broadcom custom chips, Google TPUs, AMD and Amazon Trainium account for the other reported portions of the market. Future revisions of the estimate, and any public purchasing or transition disclosures from the laboratories involved, would provide additional data points against which these reported figures can be checked.

Source: Hokanews