Trends›Specialised chips beat general ones

Specialised chips beat general ones

As AI workloads diverge (training vs inference, prefill vs decode, agents, edge), workload-specialised compute substrates take market share from general-purpose GPUs through 2030.

80 confidence, computed 23 Oct 2026

Technical
High weight
Contested

Fig. 1The evidence leans for, about 5 to one by weight, and the newest evidence points the same way.

Rests onfacts underneath

Dennard scaling has ended

What rests on thismarkets it underwrites or threatens

Underwrites

Evidencethe schedule the figure draws

The evidence, newest first

Jul 2026
for · observable Private source
Jun 2026
for · primary Broadcom AI semiconductor revenue hit $10.8B in Q2 FY26 (+143% YoY), Q3 guided to ~$16.0B (>200% YoY) on custom XPU + AI networking demand. Broadcom Q2 FY2026 results
Jun 2026
for · primary Broadcom's Q2 FY2026 AI semiconductor revenue reached $10.80 billion, up 143% YoY, powered by custom AI accelerators for hyperscalers, with Q3 FY2026 guidance of $16.0 billion AI revenue, over 200% YoY growth. Broadcom (AVGO) Q2 FY2026 Earnings: Custom AI Accelerators and AI Semiconductors Drive Growth
May 2026
for · primary Custom AI chips are outpacing NVIDIA GPU shipment growth in 2026 with TrendForce projecting 44.6% ASIC growth against 16.1% for merchant GPUs. Custom AI Chips Outpace Nvidia GPU Growth in 2026: ASIC Shipments Set to Triple GPU Rate
May 2026
for · primary Broadcom and Marvell together control roughly 95% of the ASIC co-design market; custom silicon offers up to 65% TCO advantage over GPUs for inference, driving hyperscaler adoption away from general-purpose GPUs. Custom AI ASICs: Broadcom and Marvell's Market Dominance and Market Share Shift
May 2026
against · primary NVIDIA Q1 FY27 data-center revenue ~$75B (+~92% YoY) with Blackwell 'sold out' — GPUs still absorbing AI demand far faster than custom silicon displaces them. NVIDIA Q1 FY27 data center nearly doubles (CNBC)
May 2026
against · observable Cerebras disclosed ~62% of revenue from a single UAE university customer (G42-linked) — merchant-accelerator demand still highly concentrated, not broad displacement. Cerebras 62% single-customer concentration (CNBC)
May 2026
for · observable Cerebras IPO'd May 2026 raising ~$5.5B (~$95B day-one cap) — public-market demand for a non-GPU inference substrate. Cerebras raises ~$5.5B in IPO (TechCrunch)
May 2026
for · primary Uber committed to production deployment of AWS Trainium3 on April 7, 2026 for model training and inference at 50% cost vs.
May 2026
for · primary Investment is increasingly concentrated in deployment of high-performance GPU clusters, in-house ASIC development, and next-generation data centers, with combined CapEx of top nine CSPs reaching $830 billion in 2026. North American AI Data Center Expansion Drives 2026 CapEx of Top Nine CSPs to US$830 Billion
Apr 2026
for · primary Meta announced in April 2026 commitment to deploy 1 gigawatt of custom MTIA chips co-developed with Broadcom, with MTIA 300-500 generations releasing at 6-month intervals targeting inference and recommendation workloads through 2027. Meta Deploys Four MTIA Chip Generations in Rapid Cadence (March-April 2026)
Mar 2026
for · primary Marvell's custom chip revenue hit $1.5 billion in fiscal 2026 with expected growth over 20% in fiscal 2027; overall revenue grew 42% YoY to $8.195 billion, driven by robust AI demand and custom silicon. Marvell Technology FY2026 Earnings: Custom Silicon Momentum and Data Center Growth
Feb 2026
for · observable Custom silicon now claims an accelerating share with 44.6% compound annual growth through 2033, more than doubling the 16.1% CAGR for GPU-based solutions. Custom Silicon Inflection 2026 | Introl Blog
Feb 2026
against · primary NVIDIA maintains 80% revenue share in 2026 as hyperscaler custom ASICs target only 10-15% market share despite CUDA ecosystem and TSMC allocation advantages. NVIDIA AI GPU Market Share 2026: ~80% of AI Accelerators | Silicon Analysts
Jan 2026
for · observable Morgan Stanley estimates custom ASICs reach ~25% of the AI inference market by 2026 (from <5% in 2023), ~44.6% CAGR, as inference becomes ~two-thirds of AI compute. Custom ASIC ~25% of inference by 2026 (Morgan Stanley via Oplexa)
Jan 2026
for · primary ASIC-based AI servers are expected to reach 27.8% by 2026, the highest share since 2023, with shipment growth for ASIC AI servers projected to outpace GPU-based systems. Global AI Server Shipments Forecast to Grow Over 28% YoY in 2026, with a Rising Share of ASIC-Based Systems
Jan 2026
for · primary Google's TPU v7 custom-designed units outshipped general-purpose GPUs in volume for the first time in January 2026, with over 1 million TPU units deployed throughout 2026 via Anthropic partnership. Google TPUs Outship General-Purpose GPUs in Volume for First Time (January 2026)
Nov 2025
for · observable Google moved 7th-gen Ironwood TPU (4,614 FP8 TFLOPS, 192GB HBM3E) to GA — a production-scale specialised substrate competing with GB300. Google Ironwood TPU (7th gen) GA
Oct 2025
for · primary Private source Anthropic to expand use of Google Cloud TPUs

What would move itthe measures watched

watchCustom ASIC share of AI accelerator / inference market. Rising share (esp. inference) supports the thesis; stalling near low-teens % against NVIDIA ~80% weakens it
watchBroadcom + Marvell custom XPU revenue & YoY growth. Accelerating XPU revenue + multi-GW backlog = specialised silicon scaling; deceleration = NVIDIA absorbing demand
watchHyperscaler custom-chip deployment (TPU v7, Trainium3, Maia, MTIA). Named GW-scale commitments and GA milestones move confidence up; slips/cancellations move it down
watchNVIDIA data-center GPU revenue & implied share trajectory. Continued ~90% YoY DC growth means GPUs still absorbing the wave; the thesis needs share erosion not absolute decline
watchInference vs training share of total AI compute spend. Inference majority favours workload-specialised/cheaper substrates; the bet rests on inference dominating