Version 2

NVIDIA has held or grown AI compute market share against custom ASICs as of mid-2026, reversing two years of predictions that custom silicon would erode NVIDIA's position [1]. A development from mid-June complicates the simple competitive framing: NVIDIA's NVLink Fusion technology allows hyperscalers to integrate non-NVIDIA silicon into NVIDIA-led clusters, positioning custom chip makers like Marvell as ecosystem participants rather than pure competitors [4][5]. Jensen Huang personally appeared at a Marvell keynote to make this point, and projects continued rapid growth [2][4]. Meanwhile, startups Tensordyne and DeepAdapt have each published large performance claims — 13x throughput [6] and 82% cost reduction with 33x faster inference [7] respectively — both based on unverified internal figures.

If custom silicon increasingly operates within NVIDIA's interconnect infrastructure via NVLink Fusion rather than replacing it, NVIDIA retains value from the growth of custom chips rather than losing share to them. The question for the competitive landscape shifts from whether ASICs displace NVIDIA to which parts of the AI compute stack NVIDIA continues to control.

For roughly two years, the dominant expectation in AI compute markets was that custom ASICs — built by hyperscalers and startups alike — would gradually absorb workloads from NVIDIA GPUs. Mid-2026 data has not confirmed that prediction: NVIDIA has held or grown its market share against ASICs [1], and Jensen Huang has publicly projected continued rapid growth into the following year [2]. The analytical explanation that has gained the most traction centers on software rather than silicon. SemiAnalysis states directly that AI chip competition is decided by software ecosystem depth: 100% of chip startups publish slides with impressive simulated benchmark numbers, but 99% fail in production because building a competitive software stack alongside the chip is the actual hard problem — one that NVIDIA's CUDA platform, accumulated over roughly two decades, has largely solved [3].

A mid-June development reframes the competitive picture in a significant way. NVIDIA's NVLink Fusion technology allows large hyperscalers — including Amazon — to integrate non-NVIDIA silicon into NVIDIA-led clusters, enabling coexistence rather than replacement [4]. Jensen Huang personally appeared at a Marvell keynote to explain connectivity's role in AI systems, a concrete signal that NVIDIA positions custom silicon makers as participants within its ecosystem rather than purely external competitors [4]. Marvell itself illustrates who occupies this ecosystem role: the company spent roughly $36 billion on acquisitions over the past decade, transforming from a consumer chip maker whose largest 2016 design win was a Wi-Fi chip in a children's toy into a major AI semiconductor supplier [5]. Under the NVLink Fusion model, the competitive framing of 'NVIDIA versus everyone' gives way to a picture where custom silicon plugs into NVIDIA's interconnect backbone.

The startup benchmark cycle continues alongside these structural developments. Tensordyne published a claim of 13x rack throughput versus NVIDIA's NVL72 GB300 on DeepSeek-R1 workloads, based on internal simulations with no independent validation [6]. DeepAdapt separately claims its runtime intelligence layer cuts AI operating costs by 82% and delivers 33x faster inference by shifting repetitive workloads from GPUs to standard CPUs [7] — also without third-party verification. Both announcements fit the pattern SemiAnalysis describes: large numbers on slides, simulation basis, no external confirmation [3].

Hyperscalers remain a structurally distinct case. Meta has expanded its MTIA custom silicon program to power next-generation AI workloads [8], and Amazon is entering the AI chip market [9]. Unlike independent startups, these organizations have the engineering scale and captive, predictable workloads to develop supporting software stacks alongside custom hardware. Whether their chips operate within the NVLink Fusion ecosystem or eventually provide a path to independence from NVIDIA is an open question. China's domestic AI supply chain development — covering compound semiconductors and Huawei's architecture work [10] — operates outside this ecosystem frame entirely, as a separate competitive track.