The 3.3x Annual Growth Engine Powering AI's Explosion
Global AI computing capacity is doubling every seven months. That breakneck pace translates to an approximately 3.3x yearly increase in total available computing power from AI chips across all major designers. Since 2022 this sustained multiplier has unlocked larger-scale model development and greased the path to widespread consumer adoption. NVIDIA AI chips currently own over 60% of total compute, while Google and Amazon split the remaining share across their custom TPU and Trainium families.
The data comes from Epoch AI's chip production ledger, which tracks every major accelerator from Q1 2022 through Q4 2024. The page lists units shipped, compute estimates in H100e equivalents, and price tags for each chip generation—spanning NVIDIA's A100 through B200, AMD's MI250X/MI300X/MI300A, Google's TPU v4 through v6e, Huawei's Ascend 910B/910C, and Amazon's Trainium1/Trainium2. The headline figure—3.3x yearly growth—is not a rounded guess; it is derived from the cumulative compute capacity of every chip listed across every quarter.
Chip Makers and Their Quarterly Footprint
NVIDIA's share dominates the ledger. The H100/H200 family alone posted over 790,000 units in Q3 2024, with a B200-only count of 201,198 in Q4 2024. Each quarter the numbers climb: A100 units flicker down from Q1 2022's 147,645 while H100 counts surge past 500,000 by Q2 2024. AMD's Instinct MI300X reached 80,701 units in Q2 2024, and MI300A added 10,490. Huawei's Ascend 910B and 910C shipments hold steady at roughly 100,000 and 12,500 units per quarter, respectively. Amazon's Trainium2 shipments are relatively confident above 2M total across 2024–2025, with Q2 2024 alone posting 125,000 units.
Google's TPU family expands in both variety and volume. TPU v4 units run in the hundreds of thousands each quarter, while TPU v5e consistently tops 300,000 per quarter by late 2024. The newer TPU v6e appears in smaller but growing numbers—40,118 in Q3 2024 and 173,436 in Q4. Each TPU generation carries a compute estimate in H100e equivalents, allowing the page to weight Google's contribution against NVIDIA's raw unit counts.
Why 3.3x per Year Matters
A 3.3x annual multiplier means compute capacity triples roughly every twelve months. That is the difference between training a 7-billion-parameter model and a 70-billion-parameter model without changing algorithms or data size. It is also the difference between deploying a chatbot to a few thousand users and serving millions in real time. The Epoch AI page states plainly that this growth "enables larger-scale model development and consumer adoption." The data supports the claim: by Q4 2024 the total H100e-equivalent compute listed across all chips surpasses 5.5 exaFLOPs of sustained AI work, up from roughly 1.6 exaFLOPs in Q1 2022—a bit more than 3.4x in under three years.
NVIDIA's ~60% share of total compute means the remaining 40% is a crowded field. Google's TPUs alone likely claim 15–20% of the non-NVIDIA slice, given their quarterly unit counts and H100e weights. Amazon's Trainium2 and Trainium1 make up the balance, with custom inference workloads driving much of the demand. AMD's growing MI300X and MI300A installations chip away at the periphery, but the gap between NVIDIA and the rest remains large in every quarter covered.
From Q1 2022 to Q4 2024 in Numbers
- Q1 2022: 147,645 A100 units plus a few thousand H100 early units, total compute roughly 1.6 exaFLOPs in H100e equivalents.
- Q4 2024: 681,252 H100/H200 units, 201,198 B200 units, 80,701 MI300X units, 100,000 Ascend 910B units, 387,909 TPU v5e units, and 125,000 Trainium2 units. The cumulative H100e-equivalent tally crosses 5.5 exaFLOPs.
The quarterly tables are dense, but the trend is clear: every quarter adds more total compute than the previous quarter's full-year total in many prior years. The 3.3x yearly figure smooths the step-function jumps, but the underlying data shows month‑by‑month acceleration.
Custom Chips and the Broader Landscape
Huawei's Ascend 910B and 910C shipments are noted with a methodological caveat: confidence intervals in unit counts are "relatively unprincipled due to lack of revenue model." The page urges readers not to take the numbers too literally. Still, at roughly 100,000 Ascend 910B units per quarter, Huawei contributes a steady, if uncertain, stream of compute that sits outside the NVIDIA‑Google‑Amazon triangle.
Amazon's Trainium chips are flagged as "more speculative" for early units and "relatively confident" for Trainium2 total shipments exceeding 2M across 2024–2025. The earmarking of Trainium2 with a 500-unit-per-quarter compute estimate and a $4,000 per-chip price tag means Amazon's custom AI silicon is already a multi-hundred-million-dollar line item in the broader ledger.
Looking Ahead
The Epoch AI data insights page is updated regularly. New chip revisions, new manufacturers, and revised unit estimates will shift the 3.3x figure up or down, but the underlying dynamic—rapidly expanding AI compute supply—shows no sign of reversing. Each quarterly release adds more units, newer chip generations, and finer granularity. The current headline—"Total available computing capacity from AI chips across all major designers has grown by approximately 3.3x per year since 2022"—is a snapshot of a moving target, but one that has already remade what models can do and who can access them.
The implications are straightforward: as the 3.3x yearly growth rate persists, the barrier to training ever-larger models drops, and the consumer-facing AI products that depend on them multiply. The chip makers jockey for share, but the real winner is the expanding ecosystem of applications that can draw on a compute pool that triples roughly every year.
References
- Epoch AI data insights page on AI chip production. Global AI computing capacity is doubling every 7 months. Total available computing capacity from AI chips across all major designers has grown by approximately 3.3x per year since 2022, enabling larger-scale model development and consumer adoption. NVIDIA AI chips currently account for over 60% of total compute, with Google and Amazon making up much of the remainder. https://epoch.ai/data-insights/ai-chip-production
