GPU TAM is the largest and fastest-growing semiconductor segment by absolute dollar value, driven by the AI infrastructure buildout. The base case assumes AI data center GPU demand continues growing at 30–35% annually through 2028 before moderating as the hyperscaler buildout matures, while gaming GPU recovers at a steady 8–10% cycle. Bear case assumes AI GPU demand moderates faster than expected — either because inference efficiency improvements (test-time compute, distillation) reduce GPU requirements per query, or because custom ASICs (Broadcom XPUs, Google TPUs) capture a larger share of AI compute faster than projected. Bull case assumes LLM model size and inference demand continue scaling at current pace with no efficiency ceiling in sight — requiring GPU deployments 2–3× larger than current projections through 2030.
“NVIDIA Blackwell GPU ramp driving AI data center GPU revenue to an estimated $150B+ in calendar 2026 — up from ~$47B in 2024. Every major hyperscaler (AWS, Azure, Google, Meta, Oracle) on multi-year NVIDIA GPU purchase agreements. Demand described by NVIDIA CEO Jensen Huang as "insane" with lead times stretching 12+ months for GB200 systems.”
“AI inference is now driving GPU demand independent of training. Every ChatGPT query, Copilot response, and AI API call consumes GPU inference cycles. Inference workloads are growing faster than training — NVIDIA estimates inference will surpass training as the primary GPU workload by H2 2026. This creates a structural, recurring demand floor that did not exist in prior GPU cycles.”
“Sovereign AI — governments buying NVIDIA GPU clusters to build national AI infrastructure — emerging as a new demand category. Saudi Arabia (500K H100s), UAE, France, Japan, India, and Singapore have all announced or completed significant GPU cluster purchases. Sovereign AI GPU purchases estimated at $15–20B in 2026, a category that was near-zero two years ago.”
“AMD MI350 (successor to MI300X) shipping to Microsoft Azure, Meta, and Oracle. MI350 offers 288GB HBM3E — highest memory capacity of any AI accelerator. AMD guided Instinct AI GPU revenue toward $5B+ in FY26, up from ~$3.8B in FY25. While still a fraction of NVIDIA's revenue, AMD is the only credible alternative at scale and winning incremental hyperscaler contracts.”
“Gaming GPU cycle: NVIDIA RTX 50 series (Blackwell architecture) and AMD RX 9000 series launching simultaneously in 2026 — the first overlapping next-gen gaming GPU cycle in years. AI-powered frame generation (DLSS 4, FSR 4) and on-device gaming AI features driving upgrade demand from a large installed base of RTX 20/30 series cards that skipped the RTX 40 generation.”
“US export controls on advanced AI GPUs (A100, H100, H200, B100 equivalent) to China, Russia, and certain other countries are carving out a significant portion of the addressable market. China represented ~20–25% of NVIDIA's data center revenue pre-controls. Huawei Ascend 910C is filling some of the domestic China gap, but at lower performance levels. NVIDIA estimates ~$10–15B in annual lost revenue from export restrictions.”
“Professional and workstation GPU demand recovering. NVIDIA RTX 6000 Ada and upcoming Blackwell-based professional cards seeing strong demand from VFX, CAD, and AI-assisted design workflows. Adobe, Autodesk, and Dassault Systèmes all embedding GPU-accelerated AI features requiring professional-grade VRAM (48–96GB). Professional GPU ASPs up 20%+ Y/Y.”
Training a frontier model is a one-time event, but inference is perpetual — every user query, every API call, every AI-embedded product consumes GPU cycles continuously. As AI features embed in every major software product (Office, Google Workspace, Adobe Creative Suite, Salesforce), inference GPU demand compounds with both AI adoption and usage growth simultaneously. NVIDIA estimates inference will exceed training as the primary GPU workload by H2 2026.
Global AI API calls 2026: ~10 trillion/month (OpenAI, Google, Anthropic, etc)
Each call consumes ~0.001–0.01 GPU-seconds of H100 inference compute
10T calls/month × 0.005 GPU-sec avg × $3/GPU-hr = $417M/month inference spend
Growing at 3× per year as AI adoption scales: → $4.5B/month by 2028
+ Enterprise on-premise inference deployments (non-cloud): ~40% additional
→ $80–140B cumulative incremental inference GPU TAM vs. 2024 baseline
Nations treating AI compute as strategic national infrastructure — analogous to power grids or telecommunications networks. Saudi Arabia, UAE, France, Japan, India, Malaysia, and Singapore have all announced sovereign AI GPU cluster programs, with purchases in the hundreds of thousands of H100/H200/B200 GPUs. Sovereign AI is a demand category that did not exist 24 months ago and is now estimated at $15–20B in 2026 GPU purchases alone.
Saudi Arabia: 500K H100s announced = ~$15B GPU spend
UAE + France + Japan + India + Singapore: est. ~500K GPUs combined = ~$15B
Next wave (Indonesia, Brazil, Korea, others): ~300K GPUs by 2027 = ~$9B
Total sovereign AI GPU 2026–2029 cumulative: ~$40–70B
→ This is purely incremental demand — sovereign buyers are not displacing hyperscalers
Google TPU v6, Meta MTIA 2, Amazon Trainium 3, and Microsoft Maia — all built by Broadcom — are custom AI accelerators that displace NVIDIA GPUs for specific workloads. Each hyperscaler is deploying a mix: NVIDIA GPUs for flexible, general-purpose AI and custom ASICs for specific high-volume inference workloads where fixed-function efficiency wins. The question is not whether custom ASICs displace GPUs, but how fast and for what fraction of workloads.
Custom ASIC share of hyperscaler AI compute: ~15% today → ~30–35% by 2029
Hyperscaler AI GPU spend 2029 base: ~$130B
20pp share shift from NVIDIA to custom ASICs: -$26B NVIDIA displacement
Partially offset: displaced GPU budget funds ASIC NRE and silicon (Broadcom wins)
Net NVIDIA GPU TAM impact: -$25–50B below GPU-only baseline by 2029–2030
NVIDIA RTX 50 (Blackwell) and AMD RX 9000 series launched simultaneously in early 2026, creating the first overlapping next-gen gaming GPU cycle in years. A large cohort of RTX 20/30 series owners skipped the RTX 40 generation — making them candidates for a 4–5 year upgrade. AI-powered frame generation (DLSS 4, FSR 4) dramatically changes the value proposition: users get 2–4× effective performance at the same rendering resolution. Gaming GPU market estimated to grow 12% in 2026.
Gaming discrete GPU installed base: ~300M units globally
RTX 20/30 cohort skipping RTX 40: ~120M potential upgraders
Typical GPU upgrade cycle: 3–5 years; 2026 is 4–5 years since RTX 30 peak
15% upgrade rate × 120M units × $400 avg GPU ASP = $7.2B incremental revenue
→ Split roughly 85% NVIDIA / 12% AMD / 3% Intel Arc