全球 AI 算力成長｜資料解說 | Trend Core

資料：EPOCH AI ／ 中文整理：TREND CORE

全球 AI 算力成長

從每季售出或出貨的 AI 晶片，看新增供給所代表的算力；先分清新增量，再談成長速度。

每季新增 AI 晶片供給的 H100 等效算力

圖表可左右滑動，最新季度在右側；完整數字可展開下方資料表。

每季新增 AI 晶片供給的 H100 等效算力依季度呈現新增 H100 等效算力中位估計；完整數據見圖下資料表。每季新增 AI 晶片供給的 H100 等效算力H100 等效算力｜當季新增，非累積存量｜* 資料不完整01.1百萬2.2百萬3.3百萬4.4百萬2022Q1 *：46,554 H100e2022Q1 *2022Q2：71,362 H100e2022Q3：73,961 H100e2022Q32022Q4 *：90,886 H100e2022Q4 *2023Q1：116,658 H100e2023Q12023Q2：244,637 H100e2023Q3：393,666 H100e2023Q32023Q4：533,581 H100e2024Q1：924,735.8 H100e2024Q12024Q2：1,119,391.8 H100e2024Q3：1,361,180.8 H100e2024Q32024Q4：1,903,425.8 H100e2025Q1：2,717,511.8 H100e2025Q12025Q2：3,047,098.8 H100e2025Q3：3,891,725.8 H100e2025Q32025Q4 *：1,547,338.8 H100e2025Q4 *Epoch AI ／ Josh You、Venkat Somala、Yafah Edelman、Luke Emberson｜Global AI computing capacity is doubling every 7 months｜Trend Core 重繪https://epoch.ai/data-insights/ai-chip-production｜CC BY 4.0 https://creativecommons.org/licenses/by/4.0/期間 2022-01-01 至 2025-12-31｜快照 20261019T021600187369+0000-176312511ce6｜版本 10e433e46f916e467295d944df128372c796067cc8f4fcad3d144cc618f01ec6中位估計，不是精確測量；* 資料不完整。來源列上下界不代表總和信賴區間。開啟原文CC BY 4.0

中位估計，並非精確測量。* 與斜線表示不完整，? 表示未知。

資料與原始分析：Epoch AI ／ Global AI computing capacity is doubling every 7 months · Josh You、Venkat Somala、Yafah Edelman、Luke Emberson · CC BY 4.0 · Trend Core 重繪

資料期間 2022-01-01 至 2025-12-31；快照 20261019T021600187369+0000-176312511ce6；版本 10e433e46f916e467295d944df128372c796067cc8f4fcad3d144cc618f01ec6

2025Q4，資料不完整：當季新增供給中位估計 1,547,338.8 H100e。本季未列 AMD，不代表零供給；總量僅計入有列出的估計。

圖中每一季代表該季售出或出貨的晶片所對應的算力，以總量呈現。柱子越高，表示這一季新增的紙面算力越多；這不是截至該季仍在使用的累積算力，也不是已經接電上線的算力。

這份圖使用原文附檔的歷史版本，涵蓋至 2025 年第四季。不同公司的資料起點不同，前期看到的供給結構也受資料缺漏影響。讀圖時應同時查看不完整季度標記，避免把估計缺口讀成需求突然減少。

Epoch AI 在 2026 年 1 月 9 日發表的原文中，估計自 2022 年起算力供給每年約為前一年的 3.3 倍，90% 信賴區間為 2.7 至 4.1 倍；相當於約每 7 個月翻倍，90% 信賴區間為 6 至 8 個月。這是原文的歷史估計，本頁沒有用其後更新的下載檔重新擬合，也沒有據此預測未來供給。

定義與閱讀限制

H100 等效算力，H100e
以一顆 NVIDIA H100 的峰值密集 8 位元運算能力為基準，將各型晶片的規格與估計數量換成共同尺度。這是紙面算力換算，不能直接代表實際工作負載的執行速度。
不完整季度
來源以 Incomplete 標記部分估計，例如財報季度與日曆季度未完全重合。保留數值供閱讀，但不把它當成已涵蓋整季的最終結果。
當季新增與累積
當季新增只計入該季供給；累積則將多季供給加總，還需另處理退役與折舊等問題。本圖只呈現前者。
估計上下界
原始列在有資料時保留第 5 與第 95 百分位數；缺值維持缺值。各列不確定性可能相關，不能直接相加成整體的 90% 信賴區間。
  • 這是銷售或出貨估計，不是晶片實際部署、可用算力、公司營收或現金收付。不同設計公司的證據與估計方法並不完全相同。
  • NVIDIA 資料始於 2022 年第一季，Google 始於 2022 年第四季，AMD、Amazon、Huawei 始於 2024 年第一季；未涵蓋部分不補值，這份歷史檔也不能代表此後的完整市場。
  • 規格換算未反映記憶體、軟體、網路連結與利用率，不能當成真實效能排名。
  • 本頁沒有計算折舊、退役或出貨至上線的時差，也沒有把原文成長率套成預測曲線。

從資料走向公司與專案研究

AI 機房與資本支出總覽微軟：錢花到哪裡Google：錢花到哪裡NVIDIA 資金流向
完整數據與原始估計範圍
季度總量與資料完整程度
季度當季新增 H100e 中位估計完整程度
2022Q146,554不完整：NVIDIA；未列：AMD、Amazon、Google、華為，不視為零
2022Q271,362來源未標示不完整；未列：AMD、Amazon、Google、華為，不視為零
2022Q373,961來源未標示不完整；未列：AMD、Amazon、Google、華為，不視為零
2022Q490,886不完整：Google；未列：AMD、Amazon、華為，不視為零
2023Q1116,658來源未標示不完整；未列：AMD、Amazon、華為，不視為零
2023Q2244,637來源未標示不完整；未列：AMD、Amazon、華為，不視為零
2023Q3393,666來源未標示不完整；未列：AMD、Amazon、華為，不視為零
2023Q4533,581來源未標示不完整；未列：AMD、Amazon、華為，不視為零
2024Q1924,735.8來源未標示不完整
2024Q21,119,391.8來源未標示不完整
2024Q31,361,180.8來源未標示不完整
2024Q41,903,425.8來源未標示不完整
2025Q12,717,511.8來源未標示不完整
2025Q23,047,098.8來源未標示不完整
2025Q33,891,725.8來源未標示不完整
2025Q41,547,338.8不完整：Google、NVIDIA；未列：AMD，不視為零

來源列保留數值精度，圖表與摘要適度四捨五入。上下界未知不補零，不得逐列相加當成總和信賴區間。

來源列與原始估計範圍
來源列設計商開始結束H100e 中位估計來源下界來源上界完整程度／原註記
Q1 2022 - A100NVIDIA2022-01-012022-03-31465543576360376不完整。Estimate is incomplete because it is based on Nvidia fiscal quarters beginning 1/31/2022. Estimates generated on: 01-28-2026 12:48
Q2 2022 - A100NVIDIA2022-04-012022-06-30713625466892582Estimates generated on: 01-28-2026 12:48
Q3 2022 - A100NVIDIA2022-07-012022-09-30678555218688113Estimates generated on: 01-28-2026 12:48
Q3 2022 - H100/H200NVIDIA2022-07-012022-09-30610651487204Estimates generated on: 01-28-2026 12:48
Q4 2022 - TPU v4iGoogle2022-10-012022-12-3114408752178不完整。Estimate is incomplete because it is based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 01-28-2026 12:41
Q4 2022 - TPU v4Google2022-10-012022-12-319234682312098不完整。Estimate is incomplete because it is based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 01-28-2026 12:41
Q4 2022 - TPU v5eGoogle2022-10-012022-12-31313819444693不完整。Estimate is incomplete because it is based on Broadcom fiscal quarters beginning 10/31/2022. Estimates generated on: 01-28-2026 12:41
Q4 2022 - A100NVIDIA2022-10-012022-12-31357172749146442Estimates generated on: 01-28-2026 12:48
Q4 2022 - H100/H200NVIDIA2022-10-012022-12-31324922737238332Estimates generated on: 01-28-2026 12:48
Q4 2022 - A800NVIDIA2022-10-012022-12-318865677511520Estimates generated on: 01-28-2026 12:48
Q1 2023 - TPU v4iGoogle2023-01-012023-03-316734091019Estimates generated on: 01-28-2026 12:41
Q1 2023 - TPU v4Google2023-01-012023-03-31165351180922283Estimates generated on: 01-28-2026 12:41
Q1 2023 - TPU v5eGoogle2023-01-012023-03-317764373112453Estimates generated on: 01-28-2026 12:41
Q1 2023 - A100NVIDIA2023-01-012023-03-31185311429824114Estimates generated on: 01-28-2026 12:48
Q1 2023 - H100/H200NVIDIA2023-01-012023-03-31565364779266904Estimates generated on: 01-28-2026 12:48
Q1 2023 - A800NVIDIA2023-01-012023-03-319222707511999Estimates generated on: 01-28-2026 12:48
Q1 2023 - H800NVIDIA2023-01-012023-03-31739759809252Estimates generated on: 01-28-2026 12:48
Q2 2023 - TPU v4Google2023-04-012023-06-309160515913583Estimates generated on: 01-28-2026 12:41
Q2 2023 - TPU v5eGoogle2023-04-012023-06-30237921598633787Estimates generated on: 01-28-2026 12:41
Q2 2023 - A100NVIDIA2023-04-012023-06-30202701562826426Estimates generated on: 01-28-2026 12:48
Q2 2023 - H100/H200NVIDIA2023-04-012023-06-30150868127870179142Estimates generated on: 01-28-2026 12:48
Q2 2023 - A800NVIDIA2023-04-012023-06-308272634910823Estimates generated on: 01-28-2026 12:48
Q2 2023 - H800NVIDIA2023-04-012023-06-30322752600739805Estimates generated on: 01-28-2026 12:48
Q3 2023 - TPU v4Google2023-07-012023-09-3015504532659Estimates generated on: 01-28-2026 12:41
Q3 2023 - TPU v5eGoogle2023-07-012023-09-30494383546466470Estimates generated on: 01-28-2026 12:41
Q3 2023 - A100NVIDIA2023-07-012023-09-30173111327722592Estimates generated on: 01-28-2026 12:48
Q3 2023 - H100/H200NVIDIA2023-07-012023-09-30259631219080307406Estimates generated on: 01-28-2026 12:48
Q3 2023 - A800NVIDIA2023-07-012023-09-308956690011689Estimates generated on: 01-28-2026 12:48
Q3 2023 - H800NVIDIA2023-07-012023-09-30567804574070293Estimates generated on: 01-28-2026 12:48
Q4 2023 - TPU v5eGoogle2023-10-012023-12-317703357339100662Estimates generated on: 01-28-2026 12:41
Q4 2023 - TPU v5pGoogle2023-10-012023-12-316775346810350Estimates generated on: 01-28-2026 12:41
Q4 2023 - A100NVIDIA2023-10-012023-12-31477036496226Estimates generated on: 01-28-2026 12:48
Q4 2023 - H100/H200NVIDIA2023-10-012023-12-31422172355867500558Estimates generated on: 01-28-2026 12:48
Q4 2023 - A800NVIDIA2023-10-012023-12-31286022093725Estimates generated on: 01-28-2026 12:48
Q4 2023 - H800NVIDIA2023-10-012023-12-31199711608824769Estimates generated on: 01-28-2026 12:48
Q1 2024 - TPU v5eGoogle2024-01-012024-03-318270260965108923Estimates generated on: 01-28-2026 12:41
Q1 2024 - TPU v5pGoogle2024-01-012024-03-31385502749751184Estimates generated on: 01-28-2026 12:41
Q1 2024 - H100/H200NVIDIA2024-01-012024-03-31550503466753647115Estimates generated on: 01-28-2026 12:48
Q1 2024 - H20NVIDIA2024-01-012024-03-31208071602127186Estimates generated on: 01-28-2026 12:48
FY24 Q1 - MI250XAMD2024-01-012024-03-31587739648213Date estimates were generated: 01-21-2026 16:32
FY24 Q1 - MI300XAMD2024-01-012024-03-31455143282562480Date estimates were generated: 01-21-2026 16:32
FY24 Q1 - MI300AAMD2024-01-012024-03-31350510886311Date estimates were generated: 01-21-2026 16:32
Q1 2024 Ascend 910B華為2024-01-012024-03-3135378.7527362.7545990.5Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q1 2024 Ascend 910C華為2024-01-012024-03-3196494876.7518995.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q1 2024 Trainium1/InferentiaAmazon2024-01-012024-03-3148000未知未知More speculative
Q1 2024 Trainium2Amazon2024-01-012024-03-3184250未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q2 2024 - TPU v5eGoogle2024-04-012024-06-307440650712104826Estimates generated on: 01-28-2026 12:41
Q2 2024 - TPU v5pGoogle2024-04-012024-06-30671164557891685Estimates generated on: 01-28-2026 12:41
Q2 2024 - H100/H200NVIDIA2024-04-012024-06-30648803552079761640Estimates generated on: 01-28-2026 12:48
Q2 2024 - H20NVIDIA2024-04-012024-06-30334722585343450Estimates generated on: 01-28-2026 12:48
FY24 Q2 - MI250XAMD2024-04-012024-06-3012862632380Date estimates were generated: 01-21-2026 16:32
FY24 Q2 - MI300XAMD2024-04-012024-06-3010663683501135829Date estimates were generated: 01-21-2026 16:32
FY24 Q2 - MI300AAMD2024-04-012024-06-3010395386017545Date estimates were generated: 01-21-2026 16:32
Q2 2024 Ascend 910B華為2024-04-012024-06-3035378.7527362.7545990.5Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q2 2024 Ascend 910C華為2024-04-012024-06-3096494876.7518995.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q2 2024 Trainium1/InferentiaAmazon2024-04-012024-06-3048000未知未知More speculative
Q2 2024 Trainium2Amazon2024-04-012024-06-3084250未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q3 2024 - TPU v5eGoogle2024-07-012024-09-307115242900106251Estimates generated on: 01-28-2026 12:41
Q3 2024 - TPU v5pGoogle2024-07-012024-09-307434146247107270Estimates generated on: 01-28-2026 12:41
Q3 2024 - TPU v6eGoogle2024-07-012024-09-30372191517266946Estimates generated on: 01-28-2026 12:41
Q3 2024 - H100/H200NVIDIA2024-07-012024-09-30791332675899930021Estimates generated on: 01-28-2026 12:48
Q3 2024 - H20NVIDIA2024-07-012024-09-30410503174752997Estimates generated on: 01-28-2026 12:48
FY24 Q3 - MI250XAMD2024-07-012024-09-308613071488Date estimates were generated: 01-21-2026 16:32
FY24 Q3 - MI300XAMD2024-07-012024-09-3012482999749156671Date estimates were generated: 01-21-2026 16:32
FY24 Q3 - MI300AAMD2024-07-012024-09-30431193255655194Date estimates were generated: 01-21-2026 16:32
Q3 2024 Ascend 910B華為2024-07-012024-09-3035378.7527362.7545990.5Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q3 2024 Ascend 910C華為2024-07-012024-09-3096494876.7518995.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q3 2024 Trainium1/InferentiaAmazon2024-07-012024-09-3048000未知未知More speculative
Q3 2024 Trainium2Amazon2024-07-012024-09-3084250未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q4 2024 - TPU v5eGoogle2024-10-012024-12-31598153663787974Estimates generated on: 01-28-2026 12:41
Q4 2024 - TPU v5pGoogle2024-10-012024-12-31622873981489015Estimates generated on: 01-28-2026 12:41
Q4 2024 - TPU v6eGoogle2024-10-012024-12-31160903109262224144Estimates generated on: 01-28-2026 12:41
Q4 2024 - H100/H200NVIDIA2024-10-012024-12-31681252580308796953Estimates generated on: 01-28-2026 12:48
Q4 2024 - H20NVIDIA2024-10-012024-12-31486683755363471Estimates generated on: 01-28-2026 12:48
Q4 2024 - B200NVIDIA2024-10-012024-12-31508333434276593355Estimates generated on: 01-28-2026 12:48
FY24 Q4 - MI250XAMD2024-10-012024-12-316923451101Date estimates were generated: 01-21-2026 16:32
FY24 Q4 - MI300XAMD2024-10-012024-12-31152359122474188915Date estimates were generated: 01-21-2026 16:32
FY24 Q4 - MI300AAMD2024-10-012024-12-31518393968164966Date estimates were generated: 01-21-2026 16:32
Q4 2024 Ascend 910B華為2024-10-012024-12-3135378.7527362.7545990.5Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q4 2024 Ascend 910C華為2024-10-012024-12-3196494876.7518995.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q4 2024 Trainium1/InferentiaAmazon2024-10-012024-12-3148000未知未知More speculative
Q4 2024 Trainium2Amazon2024-10-012024-12-3184250未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q1 2025 - TPU v5eGoogle2025-01-012025-03-31190241220827400Estimates generated on: 01-28-2026 12:41
Q1 2025 - TPU v5pGoogle2025-01-012025-03-31445892001769213Estimates generated on: 01-28-2026 12:41
Q1 2025 - TPU v6eGoogle2025-01-012025-03-31376006280162488240Estimates generated on: 01-28-2026 12:41
Q1 2025 - H100/H200NVIDIA2025-01-012025-03-31328883274294392144Estimates generated on: 01-28-2026 12:48
Q1 2025 - H20NVIDIA2025-01-012025-03-31541994460265766Estimates generated on: 01-28-2026 12:48
Q1 2025 - B200NVIDIA2025-01-012025-03-31131617011219221537659Estimates generated on: 01-28-2026 12:48
FY25 Q1 - MI300XAMD2025-01-012025-03-3111256782928151012Date estimates were generated: 01-21-2026 16:32
FY25 Q1 - MI300AAMD2025-01-012025-03-31452321567400Date estimates were generated: 01-21-2026 16:32
FY25 Q1 - MI325XAMD2025-01-012025-03-31431372626469226Date estimates were generated: 01-21-2026 16:32
Q1 2025 Ascend 910B華為2025-01-012025-03-3116816.511946.523883.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q1 2025 Ascend 910C華為2025-01-012025-03-31115147.2597198.5136010.75Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q1 2025 Trainium2Amazon2025-01-012025-03-31286450未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q2 2025 - TPU v5pGoogle2025-04-012025-06-3014766404524782Estimates generated on: 01-28-2026 12:41
Q2 2025 - TPU v6eGoogle2025-04-012025-06-30611825475367756958Estimates generated on: 01-28-2026 12:41
Q2 2025 - H100/H200NVIDIA2025-04-012025-06-30227846186151278409Estimates generated on: 01-28-2026 12:48
Q2 2025 - H20NVIDIA2025-04-012025-06-30228951928627021Estimates generated on: 01-28-2026 12:48
Q2 2025 - B200NVIDIA2025-04-012025-06-3011016129434511285734Estimates generated on: 01-28-2026 12:48
Q2 2025 - B300NVIDIA2025-04-012025-06-30544384461324641408Estimates generated on: 01-28-2026 12:48
FY25 Q2 - MI300XAMD2025-04-012025-06-30604504115286485Date estimates were generated: 01-21-2026 16:32
FY25 Q2 - MI325XAMD2025-04-012025-06-30353742112157668Date estimates were generated: 01-21-2026 16:32
FY25 Q2 - MI350XAMD2025-04-012025-06-309533411416212Date estimates were generated: 01-21-2026 16:32
Q2 2025 Ascend 910B華為2025-04-012025-06-3016816.511946.523883.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q2 2025 Ascend 910C華為2025-04-012025-06-30115147.2597198.5136010.75Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q2 2025 Trainium2Amazon2025-04-012025-06-30286450未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q3 2025 - TPU v6eGoogle2025-07-012025-09-30676779522439854596Estimates generated on: 01-28-2026 12:41
Q3 2025 - TPU v7Google2025-07-012025-09-3011949955837182716Estimates generated on: 01-28-2026 12:41
Q3 2025 - H100/H200NVIDIA2025-07-012025-09-30126826103574155230Estimates generated on: 01-28-2026 12:48
Q3 2025 - H20NVIDIA2025-07-012025-09-30256721683040Estimates generated on: 01-28-2026 12:48
Q3 2025 - B200NVIDIA2025-07-012025-09-308767637517151021033Estimates generated on: 01-28-2026 12:48
Q3 2025 - B300NVIDIA2025-07-012025-09-30136702211641291611710Estimates generated on: 01-28-2026 12:48
FY25 Q3 - MI300XAMD2025-07-012025-09-30605393710888954Date estimates were generated: 01-21-2026 16:32
FY25 Q3 - MI325XAMD2025-07-012025-09-30261021061947235Date estimates were generated: 01-21-2026 16:32
FY25 Q3 - MI350XAMD2025-07-012025-09-308461263241110536Date estimates were generated: 01-21-2026 16:32
FY25 Q3 - MI355XAMD2025-07-012025-09-30315032023044483Date estimates were generated: 01-21-2026 16:32
Q3 2025 Ascend 910B華為2025-07-012025-09-3016816.511946.523883.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q3 2025 Ascend 910C華為2025-07-012025-09-30115147.2597198.5136010.75Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q3 2025 Trainium2Amazon2025-07-012025-09-30387550未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
Q4 2025 - TPU v6eGoogle2025-10-012025-12-31237563181091306958不完整。Estimate is incomplete because it is based on Broadcom fiscal quarters ending 11/2/2025. Estimates generated on: 01-28-2026 12:41
Q4 2025 - TPU v7Google2025-10-012025-12-316799131769103959不完整。Estimate is incomplete because it is based on Broadcom fiscal quarters ending 11/2/2025. Estimates generated on: 01-28-2026 12:41
Q4 2025 - H100/H200NVIDIA2025-10-012025-12-31232951900528561不完整。Estimate is incomplete because it is based on Nvidia fiscal quarters ending 10/26/2025. Estimates generated on: 01-28-2026 12:48
Q4 2025 - B200NVIDIA2025-10-012025-12-31244032208983284017不完整。Estimate is incomplete because it is based on Nvidia fiscal quarters ending 10/26/2025. Estimates generated on: 01-28-2026 12:48
Q4 2025 - B300NVIDIA2025-10-012025-12-31454944387803536446不完整。Estimate is incomplete because it is based on Nvidia fiscal quarters ending 10/26/2025. Estimates generated on: 01-28-2026 12:48
Q4 2025 Ascend 910B華為2025-10-012025-12-3116816.511946.523883.25Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q4 2025 Ascend 910C華為2025-10-012025-12-31115147.2597198.5136010.75Confidence intervals in units are relatively unprincipled due to lack of revenue model. Don't take them too literally. H100e conversion incorporates some uncertainty over average chip specifications. See methodology
Q4 2025 Trainium2Amazon2025-10-012025-12-31387550未知未知Relatively confident that total Trainium2 shipments were >2M in 2024-2025
來源、方法與版本紀錄

Nvidia 自2022Q1、Google 自2022Q4，AMD、Amazon、Huawei 自2024Q1；不補缺漏。

文章附檔歷史版本，後於原文發表日；3.3倍成長率只作原文歷史估計，不對此下載檔重新擬合。

中文翻譯、編排與圖表重繪：Trend Core。依 CC BY 4.0 使用，未經 Epoch AI 審閱或背書。

原始檔涵蓋 2022-01-01 至 2025-12-31；圖表呈現期間列於圖下。

閱讀來源方法 · 下載來源資料

原始研究發表日期
2026-01-09
來源資料更新日期
2026-02-16
資料取得時間
2026-10-19T02:09:59.744963+00:00
接受版本時間
2026-10-19T02:21:14.905539+00:00
中文內容更新
2026-10-19
來源 SHA-256
315644760db099bb98df72891aedc88717782187dcb337dbcb156f43e0adf55f

晶片銷售資料集：來源更新 2026-10-14；原文發表 未提供。

完整方法：來源更新 未提供；原文發表 未提供。

資料與原始分析：Epoch AI ／ Global AI computing capacity is doubling every 7 months · Josh You、Venkat Somala、Yafah Edelman、Luke Emberson · CC BY 4.0 · Trend Core 重繪

資料期間 2022-01-01 至 2025-12-31；快照 20261019T021600187369+0000-176312511ce6；版本 10e433e46f916e467295d944df128372c796067cc8f4fcad3d144cc618f01ec6

本頁為公開資料的中文整理與視覺化，不構成投資建議。估計、預定與實際完成須分別核對。