NVIDIA vs. Custom ASICs: GPU Dominance Persists Despite Startup Performance Claims

For roughly two years, market observers expected custom ASICs to erode NVIDIA's dominant share of AI compute. Mid-2026 data suggests the opposite has occurred — NVIDIA has held or grown its position [1], and NVIDIA's CEO projects continued rapid growth [2]. SemiAnalysis attributes this to software: NVIDIA's CUDA ecosystem is the actual moat, and 99% of chip startups fail because they underestimate software depth relative to hardware specs [3]. Against this backdrop, startup Tensordyne claims 13x the rack throughput of NVIDIA's NVL72 GB300 on DeepSeek-R1 — but the figure is based on internal simulations with no independent validation [4], following the pattern SemiAnalysis describes.

Whether NVIDIA's software moat is durable determines how much of AI compute infrastructure remains captured by a single vendor. Hyperscalers — Meta and Amazon — are building custom silicon with the engineering scale to co-develop supporting software stacks, which may be structurally different from independent startups [5][6]. If SemiAnalysis is right that AI chips are fundamentally a software problem, the barrier to unseating NVIDIA is higher than any hardware benchmark suggests.

For roughly two years, the dominant prediction in semiconductor markets was that custom ASICs — designed by hyperscalers or by startups — would gradually absorb AI workloads from NVIDIA GPUs, compressing NVIDIA's market share. As of mid-2026, the data tell the opposite story: NVIDIA has held or grown its position against ASICs, reversing a widely held expectation [1]. NVIDIA's CEO has publicly projected that rapid growth will continue into the following year [2].

The explanation that has gained the most analytical traction centers on software rather than silicon. SemiAnalysis states directly that AI chip competition is fundamentally decided by software maturity: 100% of chip startups produce slides with simulated benchmark numbers showing outperformance, but 99% of custom ASIC projects fail — because building a chip and populating a slide is straightforward, while building the surrounding software ecosystem is not [3]. NVIDIA's CUDA platform, accumulated over roughly two decades, functions as the principal barrier to displacement.

This pattern is on display in the latest benchmark cycle. Tensordyne, a new inference hardware company, published a claim of 13x rack throughput versus NVIDIA's NVL72 GB300 on DeepSeek-R1 workloads [4]. The caveat is significant: the figures are based on internal simulations, not third-party or independent testing. Rohan Paul, who covered the announcement, called it a 'massive breakthrough' while acknowledging the simulation basis [4]. No independent validation has appeared.

Hyperscalers present a structurally distinct case. Meta announced an expansion of its MTIA custom silicon to power AI workloads [5], and Amazon is entering the AI chip market [6]. Unlike independent startups, large cloud providers have the engineering scale and captive, predictable workloads to develop the software stacks necessary to make custom silicon viable. The degree to which this displaces NVIDIA in hyperscaler accounts — versus shifting only specific, well-understood inference tasks — remains an open question. China's domestic AI supply chain development, including positions in compound semiconductors and Huawei's architecture work, adds a separate competitive variable operating outside the NVIDIA-versus-startup frame [7].