Cloud computing giants AWS, Google, Microsoft, Meta and OpenAI are accelerating in-house development of custom application‑specific integrated circuits, aiming to erode NVIDIA’s dominance in high‑performance AI datacentres. Industry reports highlight a projected annual growth rate of around 50% for ASIC purchases by hyperscalers, marking a strategic pivot in the AI hardware landscape.
NVIDIA’s premium-priced solutions—including Blackwell GPUs—have placed pressure on hyperscalers to secure more cost‑efficient, scalable systems. With single GPUs ranging from $70,000 to $80,000 and fully configured servers tallying up to $3 million, these companies are betting on internal design to manage costs and supply risks.
Amazon Web Services has notably moved ahead with its in‑house chips—Trainium for training and Inferentia for inference—reporting 30 – 40% greater cost efficiency compared with NVIDIA hardware. AWS is also collaborating with Marvell and Taiwan’s Alchip on next‑generation Trainium versions. Internal indications suggest AWS may deploy as many as half‑a‑million ASIC units in its data centres, an expansive scale‑up that could rival NVIDIA’s installed base.
Google, meanwhile, has scaled its TPU v6 Trillium chips, transitioning from single‑supplier to dual‑supplier design by partnering with MediaTek. With deployments reportedly hitting 100,000‑unit clusters to support Gemini 2.0 workloads, Google claims competitive cost-performance metrics relative to NVIDIA GPUs. Microsoft’s forthcoming Maia 200 chip, co‑designed with GUC using TSMC’s 3 nm process, is scheduled for commercial release in 2026.
Meta’s Meta Training and Inference Accelerator, developed alongside Broadcom, Socionext and GUC, is expected in early 2026 on TSMC’s 3 nm node, featuring HBM3e memory—another step towards self‑sufficiency in AI compute. OpenAI has also announced a proprietary training processor, with mass production anticipated at TSMC by 2026.
Market projections reflect this tectonic shift. ASICs are poised to claim between $100 billion and $130 billion of custom AI accelerator spend by 2030, with Broadcom estimating a market of $60 billion to $90 billion by 2027. Traditional ASIC powerhouses—Broadcom, Marvell, MediaTek, Alchip and GUC—are experiencing surging demand as they support hyperscaler transitions.
Despite these developments, hyperscalers continue to reserve capacity for NVIDIA chips, recognising the GPU giant’s entrenched ecosystem—especially its CUDA software stack—and the steep technical barriers to immediate elimination of GPU dependencies.
The trend resembles historical transitions in specialised compute. Just as cryptocurrency mining moved from GPUs to ASICs for lower costs and greater efficiency, hyperscalers now aim to fragment the AI compute supply chain and diversify their hardware portfolios.
TSMC stands to benefit significantly, serving as the foundry for both NVIDIA’s mass‑market GPUs and hyperscaler ASICs. Its chairman emphasises that the competition between NVIDIA and cloud‑designed chips is ultimately beneficial to TSMC, ensuring a broad customer base.
Broadcom has emerged as a frontrunner, with its ASIC and networking chipset revenues soaring 220% to $12.2 billion in 2024. Hyperscalers are investing in clusters featuring up to one million custom XPUs over open‑Ethernet networks—an architecture that places Broadcom and Marvell in strategic positions. Networking ASICs are expected to account for 15–20% of AI data‑centre silicon budgets, rising from the 5–10% range.
Revenue trends reflect these structural shifts. Marvell has secured a multi‑year AI chip deal with AWS and anticipates its AI silicon revenue jumping from $550 million in 2024 to over $2.5 billion in 2026. Broadcom, similarly, is redirecting significant investment toward hyperscaler ASIC demand.
Nevertheless, NVIDIA retains a commanding lead in AI training and general‑purpose GPU compute. Its end‑to‑end platform—from hardware to software—remains deeply embedded in the AI ecosystem. Custom ASICs, by contrast, offer task‑specific gains but lack the breadth of software compatibility that NVIDIA enables.
Analysts caution that the AI compute landscape is evolving toward a more fragmented, mixed‑architecture model combining GPUs and ASICs. Hyperscalers’ shift signals strategic recognition of rising costs, supply constraints, and performance demands. Yet, they also underscore persistent obstacles: software ecosystem maturity, long development cycles, and the complexity of large‑scale deployment.
Questions remain regarding the timeframe in which hyperscalers can meaningfully shift workloads away from NVIDIA GPUs. Industry roadmaps project new ASIC deployments through 2026–27. Analysts expect GPU market share erosion may begin toward the end of the decade, provided in-house ASICs deliver consistent performance and efficiency.
The stage is set for a multi‑year contest in datacentre compute. NVIDIA faces increasing pressure from hyperscalers building bespoke chips to optimise workloads and control supply. The next evolution of AI infrastructure may look less like a GPU‑centric world and more like a diverse ecosystem of specialised, interlocking processors.
