Test report DSG-3506 · Rev A · tested November 17, 2026

Processors & AcceleratorsDevice under test

Nvidia’s 80% Accelerator Lead Faces AMD Helios and Google’s New TPUs

Nvidia holds an estimated 80% of the AI accelerator market, while AMD’s Helios rack, Google’s new TPUs and rival custom chips target cloud buyers and hyperscalers.

Read
4 min
Words
764
Node
7nm
Operator
Grace Kim

Spec summary

  1. Nvidia holds an estimated 80% of the AI accelerator market, against 7% or less for AMD.
  2. Nvidia’s 2024 peak share was 87%, while AMD held about 5%.
  3. AMD shares jumped 13% after Oppenheimer set a $600 target, more than $100 above market price.
  4. Google added the TPU 8tx and 8i MAX to the 8t and 8i released in April, forming a four-chip eighth-generation family.
  5. SK hynix supplies more than 50% of Nvidia’s HBM, while its largest customer represented 13.35% of revenue.
Google Unveils New TPUs, AMD Pushes Helios Into Nvidia's Gap
Fig. AGoogle Unveils New TPUs, AMD Pushes Helios Into Nvidia's Gap — monicore / pixabay

By Grace Kim

Nvidia holds an estimated 80% of the artificial intelligence accelerator market, against 7% or less for AMD, according to TrendForce and IDC data based on data-center revenue. The estimate includes chips, boards, servers and racks; Nvidia's share peaked at 87% in 2024, when AMD stood at about 5%.

The market opening follows an affair involving Nvidia Chief Executive Jensen Huang and Alice Huang, who is suspected of spying for China, leaving Huang's leadership under strain. Rivals are expected to move quickly. Hyperscalers—the operators of large-scale computing infrastructure—are likely to step up sales campaigns for custom chips designed to replace Nvidia graphics processing units (GPUs).

What does this change for infrastructure buyers?

AMD is drawing the most attention as an alternative to Nvidia. Long regarded as the perennial No. 2 in graphics processing units, the company began selling Helios in earnest as the crisis at Nvidia unfolded.

The server rack is built around AMD's latest CPUs and GPUs. Oppenheimer, an investment bank, set a $600 price target, more than $100 above the market price; AMD shares jumped 13% on the 9th, local time.

The timing gives AMD a wider route into agent infrastructure. AI agents, software assistants that perform tasks on their own, run on server central processing units (CPUs), a market AMD leads. AMD has gained ground in the agent era because agent infrastructure relies on those CPUs. The accelerator gap remains wide.

Memory demand provides a counterweight. SK hynix, Nvidia's key high-bandwidth memory (HBM) supplier, provides more than 50% of the HBM used by Nvidia. Its half-year report says its largest customer represented 13.35% of revenue.

Some industry participants call a revenue-hit reading overblown. An industry official said: “HBM is in such short supply that we cannot make enough to sell, and the customer base is already diversified.” The official added that a shift in AI accelerator competition would not translate directly into weaker memory demand.

Tighter export restrictions on China following the Jensen Huang affair could impede SK hynix's plan to raise output at its Dalian plant. SK hynix had intended to use earnings from that operation to invest in a New York state plant run by Solidigm, a subsidiary of its affiliate.

Google, the front-runner in custom AI chips, unveiled the TPU 8tx and TPU 8i MAX on the same day as AMD's share jump. The additions align Google's portfolio with the shift toward AI agents.

Google developed its tensor processing unit (TPU) line to reduce dependence on Nvidia GPUs. The chip is less versatile than a GPU but ahead on inference performance and cost efficiency.

The two new devices join the training-focused 8t and inference-focused 8i released in April, bringing the eighth-generation TPU family to four chips. With Nvidia shaken, Google is expected to use the portfolio in a sales push to cloud companies that lease data-center capacity and have retained confidence in Nvidia GPUs.

Google also worked with AMD, a company it had treated as a rival, on developing the chips, turning the two into allies.

Other custom-chip programs are already moving:

  • Microsoft unveiled Maia 300 last month, its third-generation AI accelerator aimed at countering GPUs, and is preparing a successor to the in-house Cobalt CPU.
  • Amazon Web Services already uses the Trainium AI accelerator and Graviton CPU.
  • Meta released MTIA, a custom chip for AI inference.
  • OpenAI unveiled Jalapeño in June, an inference-only chip co-developed with Broadcom.
  • Anthropic has begun developing its own chip.

The effect on South Korea's memory suppliers is uneven. Samsung Electronics (005930.KS) and SK hynix (000660.KS) are the main HBM suppliers. Some analysts see a benefit for Samsung as accelerator competition broadens. Its half-year report names Alphabet, Amazon, Apple, Hong Kong-based Techtronic and Supreme Electronics as its five largest customers. Samsung has also broadened its AI customer base through deeper cooperation with OpenAI.

Samsung and SK hynix were already adapting before the Jensen Huang affair, limiting expected disruption. Samsung unveiled zHBM in August, while SK hynix disclosed three-dimensional stacked dynamic random access memory last month. Both technologies stack memory on top of a GPU and can serve AI accelerators tailored to individual customers.

For infrastructure buyers, the expected change is a broader supplier set, not a settled shift in market control. AMD's Helios rack and Google's four-chip TPU family add options, but Nvidia still holds an estimated 80% of the market, against 7% or less for AMD.

via en.sedaily.com (Source)

Filed under

  • nvidia
  • amd-helios
  • google-tpu
  • ai-accelerators
  • hbm
Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering regulation, policy and their impact on hardware layer of artificial intelligence at Die Signal.

665 articles

Same lot · LOT-362D

« Previous articleNext article »