AI output is already in real use across our daily work, and the next step I propose is investing in hardware and an AI video production line: the video pipeline needs USD 4,600 (NT$148,212) over three gated stages, and the only decision that costs money now is USD 400; the hardware side follows a step-by-step upgrade ladder whose near-term rungs total about NT$17,600 to 29,100 including an optional monitor, with memory already narrowed to a specific NT$5,100 kit. AI 產出已經實際用在我們的日常工作上，下一步我提議投資硬體與 AI 影片產線：影片產線三階段合計 USD 4,600（NT$148,212），現在唯一要花錢的決定是核准 USD 400；硬體端走一條逐階升級的路線圖，近期各階合計約 NT$17,600 至 29,100（含選配螢幕），其中記憶體已挑好定稿品項 NT$5,100。
I measured everything below on 2026-08-13 with standard system queries and nvidia-smi, on the machine itself [S8]. Two items cannot be read from software and are flagged.下表全部是我在 2026-08-13 用系統查詢與 nvidia-smi 在這台機器上實測的結果 [S8]。有兩項軟體讀不到，已加標記。
| Component零件 | Measured state實測現況 | Verdict升級判斷 | Why理由 |
|---|---|---|---|
| CPUIntel i5-10400F處理器Intel i5-10400F | 6 cores / 12 threads, 10th gen6 核 12 緒，第 10 代 | KEEP暫不動 | Oldest part of the platform, but no data shows it maxed out today. A meaningful CPU upgrade forces a motherboard swap too.全機最舊的一環，但目前沒有實測數據顯示它滿載卡住。要換更強的 CPU 就得連主機板一起換。 |
| MotherboardGigabyte B460M DS3H V2主機板技嘉 B460M DS3H V2 | B460 chipset, BIOS F1 dated 2020-12, DDR4 platformB460 晶片組，BIOS F1 停在 2020 年 12 月，DDR4 平台 | KEEP暫不動 | Entry chipset, upgrade path capped at 10th/11th gen. New builds in 2026 are DDR5, which is why memory bought for this board will never migrate.入門晶片組，升級路線最高只到 10 或 11 代。2026 年的新機一律 DDR5 平台，所以買給這塊板的記憶體永遠搬不過去。 |
| Memory2×8GB DDR4-2666記憶體2×8GB DDR4-2666 | 16GB total, 74% to 77% used at light load (3 samples). 2 of 4 slots free, board max 64GB. Not yet expanded; nothing has been bought.共 16GB，輕度使用實測已用 74% 至 77%（三次取樣）。四槽用二，主機板上限 64GB。尚未擴充，也還沒買任何新條。 | UPGRADE NOW本輪升級 | The one upgrade backed by live measurements: only 3.6 to 4.2GB headroom before running any AI workload. Two free slots mean adding sticks, no discarding. Ladder rung 1 in section 05.唯一有實測數據支撐的升級：還沒跑 AI 工作就只剩 3.6 至 4.2GB 緩衝。有兩個空槽可以直接加，不必丟舊的。見第 05 節第一階。 |
| Graphics cardRTX 4060 Ti 8GB顯示卡RTX 4060 Ti 8GB | Newest part in the machine, driver healthy. I bought it myself at NT$12,000.全機最新的零件，驅動正常。我自費 NT$12,000 購入。 | NO COMPANY SPEND本輪不需公司出錢 | I already covered this part out of my own pocket, so no graphics-card money is requested this round. When an upgrade to 16GB would make sense is laid out in section 05, rung 4.這一項我已經自己出錢補上，所以這一輪不向公司申請任何顯卡費用。什麼時候才值得升到 16GB，寫在第 05 節第四階。 |
| System drive C:Kingston NVMe 500GB系統碟 C:Kingston NVMe 500GB | 296.8GB free, healthy剩 296.8GB，健康 | OK正常 | Enough free space to host the hot AI models as a stopgap. See section 03.剩餘空間足以先接住常用 AI 模型當過渡方案。見第 03 節。 |
| Drive D:SATA SSD 240GBD 槽SATA SSD 240GB | 19.1GB free, 92% full剩 19.1GB，已用 92% | CLEAN NOW立即清理 | Already full today. Any new download can fail on space. Costs nothing to fix, so it ranks above every purchase.現在就已經滿了，任何新下載隨時可能因空間不足失敗。清理不用錢，優先度排在所有採購之前。 |
| Drive F:Toshiba external USB 2TBF 槽Toshiba 外接 USB 2TB | 1,376GB free, healthy, currently holds the AI models剩 1,376GB，健康，AI 模型目前放在這裡 | MOVE MODELS立即搬移 | USB interface throttles every model load, today, regardless of any purchase. This is section 03.USB 介面拖慢每一次模型載入，這件事今天就在發生，跟買不買新東西無關。詳見第 03 節。 |
| Power supply電源供應器 | Wattage unknown, case must be opened查不到，需開機殼瓦數 unknown, case must be opened查不到，需開機殼 | VERIFY需開殼確認 | No software can read the rated wattage, it is only printed on the unit. Must be confirmed before adding drives or ever changing the GPU.任何軟體都讀不到額定瓦數，只印在機身標籤上。加裝硬碟或未來換卡前必須先確認這個數字。 |
| M.2 slotsM.2 插槽 | Total count unknown, manual or visual check needed查不到，需查手冊或開殼, 1 known in use總數 unknown, manual or visual check needed查不到，需查手冊或開殼，已知用掉 1 個 | VERIFY需開殼確認 | Determines whether the new NVMe drive in section 05 can be installed internally. Zero-cost check, blocks that purchase until done.決定第 05 節的新 NVMe 能不能裝進機內。確認不用錢，但確認前該筆採購先不動。 |
| OS作業系統 | Windows 11 Pro, build 26200Windows 11 專業版，組建 26200 | OK正常 | Measured as full Pro edition. Earlier notes conflicted on the edition; the discrepancy is logged in the open-items list, section 09.實測為完整專業版。先前紀錄對版本說法不一，這筆差異已列入第 09 節的待確認清單。 |
Our AI models currently live on an external USB hard drive, and every single model load pays a USB-speed penalty. Moving the frequently used models onto the internal NVMe system drive costs NT$0 and can be done today. 我們的 AI 模型現在放在外接 USB 硬碟上，每一次載入模型都在付 USB 速度的代價。把常用模型搬進內接 NVMe 系統碟，成本是 NT$0，今天就能做。
The C: drive has 296.8GB free, enough for the working set of models. Loading a large model file from NVMe versus a USB mechanical drive differs by minutes per load, and that cost repeats on every model switch [S1]. This slowdown exists now and is unrelated to any purchase decision, which is why I put it before all the spending sections.C 槽還有 296.8GB 可用，足以放進常用的那批模型。從 NVMe 載入大型模型檔與從 USB 機械硬碟載入，每次差距以分鐘計，而且每次切換模型都要再付一次 [S1]。這個拖慢現在就存在，跟任何採購決定無關，所以我把它放在所有花錢的章節之前。
Memory prices rose 3 to 4 times between mid-2025 and early 2026, and the market is now in a pullback: DDR4 spot fell 21.64% in April 2026 alone. I read this as a relative low after the spike, a window that will not necessarily stay open, and I am showing the pullback numbers here on purpose rather than hiding them. 記憶體價格從 2025 年中到 2026 年初漲了 3 到 4 倍，目前正處於回檔：DDR4 現貨價 2026 年 4 月單月就跌了 21.64%。我的判讀是現在處於暴漲後的相對低點，這個窗口未必會一直開著。回檔數字我刻意放在這裡，一起看才站得住。
The three memory makers moved wafer capacity to HBM chips for AI servers, and one HBM chip consumes roughly the capacity of three regular DDR5 chips, so consumer memory supply shrank structurally. Their 2026 capacity is already fully booked [S1][S2].三大原廠把晶圓產能轉去做 AI 伺服器用的 HBM，一顆 HBM 消耗的產能約等於三顆一般 DDR5，消費級供給因此結構性萎縮。原廠 2026 年產能早已被預訂一空 [S1][S2]。
Supporting "buy in the window": H1 2026 is still forecast to rise 5% to 10%, one analyst camp expects high prices to persist into late 2026 or 2027, and European DDR5 retail prices even after their February pullback remain over 3 times the autumn 2025 level [S1][S2]. Against it: if the H2 supply gap really closes to 1% to 2%, waiting could cost less than buying now. Both camps are forecasts; neither is proven. The timeline of the two headline numbers, stated plainly: the 5 to 10% is TrendForce's forecast for H1 2026 as a whole, itself already a convergence down from the earlier runaway; the 21.64% is the actual spot print for the single month of April. One is a half-year forecast, the other one month of reality, and reality pulled back faster and deeper than the forecast, which is what my "the pullback is real" reading rests on. Also stated plainly: the monthly spot-index data this report obtained ends at April, and the freshest memory price signal after that is the July 25 retail quote in section 05. My recommendation prices this in by keeping the memory buy at the minimum viable size and re-checking the quote on purchase day.支持「在窗口內買」的一面：2026 上半年仍預估漲 5% 至 10%，一派分析認為高價會延續到 2026 年底甚至 2027 年，歐洲通路 DDR5 就算經過 2 月回檔，價格仍是 2025 年秋季的 3 倍以上 [S1][S2]。反面：若下半年供需缺口真的收斂到 1% 至 2%，等幾個月可能比現在買更便宜。兩派都只是預估，都還沒被證實。兩個主數字的時間軸也照實講：5% 至 10% 是 TrendForce 對 2026 整個上半年的預估漲幅，而且本身就是從先前暴漲收斂下來的預估；21.64% 是四月單月實際走出來的現貨跌幅。一個是對半年的預估，一個是單月的實況，實況的回檔跑得比預估更快更深，我判讀「回檔是真的」靠的就是這個。同樣照實講：本報告取得的現貨指數月度數據到四月為止，記憶體這一項四月之後最新的價格訊號，是第 05 節那筆 7 月 25 日的零售報價。我的處理方式是把記憶體採購壓在最小可用容量，並在採購當天重看一次報價，把這個不確定性納進決策，不去賭單邊。
In pullback after a 3 to 4× run. Relative low, window uncertain. This is the item the timing argument applies to.暴漲 3 到 4 倍後正在回檔，屬相對低點，窗口長短不確定。時點論證主要就是講這一項。
NAND contract prices forecast up 70% to 75% in Q2 2026 and unlikely to fall before 2027, so later is likely more expensive, and today's price is already historically high. Both facts stated [S9].NAND 合約價 2026 第二季預估季增 70% 至 75%，業界預估 2027 年前難回落，所以越晚買越貴的機率高，同時現在的單價也已在歷史高位。兩面都照列 [S9]。
Worst premium of any category, 50% to 122% over list (RTX 5090 went from NT$71,990 to NT$159,990), and retail widely forces bundle deals. We skip this category this round [S4].全品項溢價最重，比原價貴 50% 至 122%（RTX 5090 從 NT$71,990 喊到 NT$159,990），且通路普遍強制綁組機。這一輪整個跳過 [S4]。
I planned the upgrades as a ladder with numbered rungs, where each rung states what is bought, what capability it unlocks, whether it migrates to the future company machine, the condition for moving up, and the cost of skipping it. Honesty first: this machine is a DDR4 platform and any company machine built in 2026 will be DDR5, so memory bought now stays on this transition machine, priced accordingly as transition spend. Drives, monitor, and any future graphics card all migrate.我把升級規劃成一條有編號的階梯，每一階都寫清楚買什麼、解鎖什麼能力、將來能不能移到公司電腦、達成什麼才進下一階、先不做的代價是什麼。先講誠實的那句：這台是 DDR4 平台，2026 年組的公司新機一律 DDR5，所以現在買的記憶體會留在這台過渡機上，我照實把它列為過渡期花費。硬碟、螢幕、以及未來若購入的顯示卡則全部會跟著走。
| Rung階 | Item品項 | Cost (NT$)金額（NT$） | Installed on裝在哪 | Migrates可移轉 |
|---|---|---|---|---|
| 0第 0 階 | Clean drive D:, move models off USB清理 D 槽、模型搬離 USB | 00 | Transition machine過渡機 | N/A不適用 |
| 1第 1 階 | DDR4 memory kit, 2×8GBDDR4 記憶體套條 2×8GB | 5,100quoted 2026-07-255,100查價 2026-07-25 | Transition machine過渡機 | NO否 |
| 2第 2 階 | Internal NVMe drive (company-data drive)內接 NVMe（公司資料專用碟） | 2,500–5,0001TB band, 2026-08-132,500–5,0001TB 區間，2026-08-13 | Transition machine, later pulled into the company build過渡機，之後整碟拔進公司機 | YES是 |
| 3第 3 階 | 27-inch 2K pivot monitor (optional)27 吋 2K 可旋轉直立螢幕（選配） | 10,000–19,000band price, 2026-08-1310,000–19,000區間價，2026-08-13 | Desk, serves both machines桌面，兩台機器都用得到 | YES是 |
| 4第 4 階 | 16GB graphics card (no purchase this round)16GB 顯示卡（本輪不買） | 0this round0本輪 | Whichever machine runs the image line when conditions are met條件成立時，裝在跑出圖線的那台 | YES是 |
| 5第 5 階 | Company-owned build (DDR5), endpoint公司自己的電腦（DDR5），終點 | separate proposal另案提出 | Company machine公司機 | ENDPOINT終點 |
Item, already narrowed to a specific kit: Kingston FURY Beast DDR4 3200, 16GB (8G×2), model KF432C16BBK2/16, NT$5,100 at a Taiwan retailer with lifetime warranty, quoted 2026-07-25 quote is 3 weeks old, re-confirm before ordering查價已三週，下單前需重新確認. It has not been ordered; I want the company's go-ahead first.品項已挑好定稿：Kingston FURY Beast DDR4 3200 16G（8G×2），型號 KF432C16BBK2/16，順發線上購物 NT$5,100，終身保固，查價日 2026-07-25 quote is 3 weeks old, re-confirm before ordering查價已三週，下單前需重新確認。我還沒下單，想先問過公司。
I already paid NT$12,000 out of my own pocket for the RTX 4060 Ti 8GB now in the machine, so this round asks the company for zero graphics-card money. This rung states what a dedicated verification of local AI video generation (completed 2026-08-13, first-party sources throughout) concluded, and where a 16GB card genuinely earns its place.機器裡這張 RTX 4060 Ti 8GB 是我自費 NT$12,000 買的，所以這一輪不向公司申請任何顯卡費用。這一階要講的是「本機 AI 影片生成」專項查證（2026-08-13 完成，全程一手來源）得出的結論，以及 16GB 的卡真正值得買在哪裡。
The part of my original judgment that verification confirmed: 8GB is genuinely not enough for AI video. That call was correct: on 8GB the community consensus caps out at 480p short clips on small models, and quantization cannot stretch an 8GB card into a 24GB one; past a point the output degrades below what a smaller model does natively [S11].查證確認了我原本判斷對的那一半：8GB 做 AI 影片確實不夠。這個判斷正確：8GB 上社群共識的天花板是小模型的 480p 短片段，而且量化沒辦法把 8GB 的卡硬撐成 24GB 用，壓過頭之後產出會退化到比原本就小的模型還差 [S11]。
The part that does not hold, said plainly: buying a 16GB or 32GB card fixes "cannot run", and leaves "cannot deliver publishable quality" untouched. The leaderboard's Elo is a match score: two clips go head to head in front of viewers, the better one gets picked, winners gain points and losers lose them, and a wider gap means more lopsided matchups. On the 2026-08-13 blind-test leaderboard, Seedance 2.0 at 720p scores Elo 1,223 while the highest-scoring open model a single consumer card can actually run (LTX-2.3 Fast) scores 979; a 244-point gap means the commercial model wins roughly 80% of blind comparisons. The only open model that beats it (MiniMax-H3, 1,239) is 33B with an official 4-GPU reference deployment, out of reach for any single card. Seedance 2.5 is not on the board yet, so the real gap is larger, never smaller [S11]. The money would buy a pipeline whose output still loses blind tests to a NT$37.2 cloud clip.不成立的那一半，也照實講：買 16GB 或 32GB 的卡解決的是「跑不動」，「做不出可對外發布的品質」這件事沒有被解決。先講 Elo 是什麼：它是盲測榜用的比賽積分，兩支影片一對一放給人看、挑出比較好的那支，贏的加分、輸的扣分，累積成每個模型的分數，差距越大代表對決時輸面越大。2026-08-13 的盲測排行榜上，Seedance 2.0 的 720p 是 Elo 1,223，單張消費級卡真正跑得動的最高分開源模型（LTX-2.3 Fast）只有 979，差 244 分約等於盲測八成會輸。榜上唯一贏過它的開源模型（MiniMax-H3，1,239）是 33B、官方參考部署要四張 GPU，任何單卡都進不去。Seedance 2.5 還沒上榜，所以實際差距只會更大 [S11]。這筆錢買到的產線，產出在盲測裡仍會輸給一支 NT$37.2 的雲端片段。
On cost, said straight: counting electricity alone, one 5-second generation on the local machine costs about NT$0.55 against NT$37.2 on the cloud [S11] electricity inputs: power rate not first-hand verified, 5 min/clip is an estimate電費輸入值：電價未一手驗證，每支 5 分鐘為推估, so per generation the local machine is far cheaper, and staying on the cloud was never about saving money. The reason is the paragraph above: a per-clip cost comparison only means something when both sides produce the same product, and the blind-test gap says the local clip does not reach deliverable quality. Money saved on footage that cannot ship is not savings. So video stays on the cloud pipeline in section 06, pay-per-output, as this round's production line.成本這一面照實講：只算電費，本機生成一次 5 秒片段約 NT$0.55，雲端是 NT$37.2 [S11] electricity inputs: power rate not first-hand verified, 5 min/clip is an estimate電費輸入值：電價未一手驗證，每支 5 分鐘為推估，單看每次生成的成本本機便宜得多，所以把影片留在雲端，理由從頭到尾都與省錢無關。真正的理由是上一段：每支成本要能拿來比，前提是兩邊做出來的是同一種產品，而盲測差距說明本機那支達不到可交件的品質，省在不能交件的片段上的錢等於沒省。所以影片留在第 06 節的雲端產線，按產出計費，這就是本輪的主力產線。
Where a 16GB card genuinely earns its place: image generation and training the company's own style. The rented online image services we lean on have daily caps, frequently block outputs on content review, and keep the style assets on their side rather than ours. Running the image line locally removes all three at once; the tool stays ComfyUI, only the underlying model changes, and full-size image models (FLUX.1 Dev class) list 16GB minimum, which 8GB cannot reach. One wording I want to be precise about: "training" here means teaching an existing full-size image model our own style, known as LoRA fine-tuning, which a 16GB card handles (its exact training VRAM is unverified, usually above inference未查證，通常高於推論); training a large model from scratch is out of reach for 8GB and 16GB alike and is in nobody's plan. These local advantages hold today for images: no service-rule changes can reach us, assets stay in our hands, no per-day caps, failed generations cost nothing, and client material never leaves the machine. For video the same advantages must wait for the ecosystem to mature; in the meantime the local card still helps as a support tool for storyboard tests, animating approved stills, and LoRA-consistent brand elements. This purchase sits at a later rung, after three conditions hold: premium back near list price, retail no longer forcing bundles, and the image-line need confirmed.16GB 的卡真正值得買在哪裡：出圖，加上訓練公司自己的風格。我們現在依賴的線上出圖服務有每日上限、常因內容審查擋圖、風格資產也留在服務商那邊，不在我們手上。出圖線收回本機，這三個限制同時消失；工具照樣是 ComfyUI，只換底下的模型，而完整版製圖模型（FLUX.1 Dev 級）門檻列 16GB 起，8GB 到不了。這裡有一個用詞我要講精準：所謂訓練，指的是在現成的完整版製圖模型上訓練出公司自己的風格，也就是 LoRA 微調，16GB 的卡做得到這件事（訓練所需的確切 VRAM unverified, usually above inference未查證，通常高於推論）；從零訓練一個大模型，8GB 跟 16GB 都做不到，也完全不在規劃內。本機方案的優勢在出圖上現在就成立：服務商改規則影響不到我們、資產在自己手上、沒有每日上限、生成失敗不用付費、客戶素材不出這台機器。同樣的優勢用在影片上則要等生態成熟；在那之前，本機的卡仍可當輔助工具，做分鏡試打、把定稿靜圖做微動、用 LoRA 做高度一致的品牌元素。這筆採購放在階梯較後面，等三個條件同時成立：溢價回落到接近原價、通路不再強制綁機、出圖線的需求確認成立。
| Model size模型規模 | Inference (run only)只跑（推論） | QLoRA fine-tuneQLoRA 微調 |
|---|---|---|
| 7B7B | 6–8 GB | ~12 GB |
| 13B13B | 10–12 GB | ~20 GB |
| 32B32B | 20–24 GB | ~44 GB |
| 70B70B | 40–48 GB | ~88 GB |
This table covers text AI (language models), a different family from the video and image models this rung discusses; it sits here only as a general graphics-memory reference. 4-bit quantized, 8K context [S1]. Video memory cannot be substituted by system RAM or by pairing two cards; a model that does not fit runs an order of magnitude slower. LoRA and full fine-tune figures for 13B+ are not found in verifiable sources查無可引用數字.此表講的是文字型 AI（語言模型），跟本階談的影片與出圖模型是不同類，放在這裡只作顯示記憶體的一般參考。4-bit 量化、8K 上下文 [S1]。顯示記憶體沒辦法用系統記憶體或多插一張卡湊數，模型放不進去時速度會掉一個數量級。13B 以上的 LoRA 與全參數微調數字 not found in verifiable sources查無可引用數字。
The pipeline runs on Seedance 2.5, which charges per generated output (USD 10.70 per million tokens without video input; prompts and reference images are free; a token is the billing unit here, think of it as one small piece of the finished video, so longer or sharper video means more pieces, and converted it works out to about USD 0.2312 per second at 720p, the per-second rate behind every 720p figure below) [S5][S6]. USD figures convert at 1 USD = 32.22 TWD, rate dated 2026-08-11 [S10]. One thing to connect first, because the company has heard a much smaller number from me before: in May's AI-ad proposal I said no card purchase was needed because cloud image generation rents for about US$35 a month. That figure was and still is about image generation; the USD 4,600 here funds the video pipeline, billed per generation, and the two are separate line items, a US$35 image subscription cannot produce this volume or quality of video (that subscription shows as never activated in my records; its status sits in the section 09 open-items list). Another thing said up front: I checked the official BytePlus pricing pages on 2026-08-13 and found no Seedance 2.5 entry, so all prices here come from third-party reporting, cross-checked against subscription point costs official pricing page has no entry官方定價頁查無條目.產線用 Seedance 2.5，按產出計費（不含影片輸入 USD 10.70 每百萬 tokens，提示詞與參考圖免費；token 是這類服務的計價單位，可以想成組成影片的一小塊碎片，影片越長、畫質越高，碎片就越多，換算下來 720p 約每秒 USD 0.2312，本節 720p 的數字都是用這個每秒價算的）[S5][S6]。美金換算採 1 USD = 32.22 TWD，匯率取價日 2026-08-11 [S10]。先接一句公司以前聽我講過的數字，免得對不上：五月的 AI 廣告提案我說過不用買卡，出圖租雲端服務、月費約 US$35 就能做。那句講的是出圖，現在也還是出圖的方案；這份的 USD 4,600 是影片產線的錢，按每次生成計費，兩筆是兩回事，US$35 的出圖訂閱生不出這個量與畫質的影片（那筆訂閱在我的紀錄裡停在未開通，實際狀態列在第 09 節待確認清單）。另外先講明一件事：我在 2026-08-13 實查 BytePlus 官方定價頁，查無 Seedance 2.5 條目，本節所有價格來自第三方轉述，已用訂閱點數交叉驗算 official pricing page has no entry官方定價頁查無條目。
| Stage階段 | Budget預算 | Output產出 | What it verifies驗證什麼 | Gate to next stage進下一階段的條件 |
|---|---|---|---|---|
| 1 · Template discoverymonth 1一 · 模板探索第 1 個月 | USD 400NT$12,888USD 400NT$12,888 | 28 to 32 test clips of 10 to 15s (720p), expecting 7 to 8 usable28 至 32 支 10 至 15 秒測試片段（720p），預期 7 至 8 支可用 | Whether template hit-rate can reach 40%; whether 3 reusable style templates emerge; whether the real per-second price lands inside the quoted band模板成功率能否拉到 40%；能否寫出 3 套可重複套用的風格模板；實際每秒單價是否落在轉述區間內 | At least 2 templates hit 40% usable, and actual per-second cost stays within +15% of USD 0.2312至少 2 套模板達 40% 可用率，且實際每秒成本不超過 USD 0.2312 的 +15% |
| 2 · Pipeline proofmonths 2–3二 · 產線驗證第 2 至 3 個月 | USD 1,200NT$38,664USD 1,200NT$38,664 | 8 full 90s videos (720p) plus 1 at 1080p for comparison; 5 to 6 expected deployable8 支 90 秒成片（720p）加 1 支 1080p 對照，預期 5 至 6 支可投放 | Usable-rate reaching 50%; whether real 1080p pricing lands in the assumed band; whether output runs directly on the ad account可用率能否升到 50%；1080p 實際計價是否落在假設區間；成片能否直接上廣告帳戶 | All three: usable rate ≥50%; per-second cost down ≥25% vs stage 1 (stage 1 clips run 10 to 15s, stage 2 videos 90s, so only the per-second figure compares); and ≥3 videos through real ad tests with performance data三項同時成立：可用率 50% 以上、每秒成本較階段一降 25% 以上（階段一是 10 至 15 秒片段、階段二是 90 秒成片，換算成每秒才能比）、至少 3 支完成實際投放並取得成效數據 |
| 3 · Scalemonths 4–6三 · 規模化第 4 至 6 個月 | USD 3,000NT$96,660USD 3,000NT$96,660 | 8 to 12 videos per month, 24 to 36 total, 15 to 22 deployable每月 8 至 12 支，共 24 至 36 支，預期 15 至 22 支可投放 | Unit model cost dropping below USD 60; stable monthly capacity; ad performance holding the guardrail單支模型成本能否降到 USD 60 以下；月產能是否穩定；投放成效是否守住護欄線 | Stop rules: any month above stage-2 unit cost pauses scaling; CTR or ROAS below 80% of the current-creative baseline pauses new volume停損規則：任一個月單支成本高於階段二基線即暫停擴量；CTR 或 ROAS 低於現有素材基線的 80% 即暫停新增產量 |
| Total, 6 months六個月合計 | USD 4,600NT$148,212USD 4,600NT$148,212 | 33 to 45 full 90s videos targeted, count includes stage 2's 1080p comparison video目標 33 至 45 支 90 秒成片，含階段二的 1080p 對照支 | 20 to 28 expected deployable預期 20 至 28 支可投放 | Any missed gate stops the spend at that stage任一關卡沒過，錢就停在該階段 |
Unit economics behind the table: a 90-second video needs about 15 shots of 6 seconds, and with a 3-to-5 generations-per-keeper retry assumption, one 720p video costs USD 62 to 104 (NT$1,998 to 3,351) at the verified third-party rate. The 1080p figure of USD 95 to 248 rests on an interpolated per-second price of USD 0.35 to 0.55 1080p official price unpublished, estimate only1080p 官方未公布，此為推估, and the retry multiple itself is an assumption and the single most sensitive number in the plan: moving it from 3 to 5 raises unit cost by 67% [S5][S6]. Stage 1's USD 400 follows the same arithmetic: the 28 to 32 clips are the kept test clips, each costed at the 15-second upper bound times the same 3-to-5 retry multiple, so one kept clip means 45 to 75 seconds actually generated, about USD 10.4 to 17.3; USD 400 covers 23 to 38 such clips and the plan takes the middle band of 28 to 32, with the retry spend already inside the 400, and the stage stops when the ceiling is spent.表格背後的單支算法：一支 90 秒影片約需 15 顆 6 秒鏡頭，採 3 至 5 次生成挑 1 支的重試假設，720p 單支成本為 USD 62 至 104（NT$1,998 至 3,351），用的是已交叉驗算的第三方轉述價。1080p 的 USD 95 至 248 建立在每秒 USD 0.35 至 0.55 的內插推估上 1080p official price unpublished, estimate only1080p 官方未公布，此為推估，而重試倍數本身是假設，也是整個方案最敏感的參數：倍數從 3 變 5，單支成本就漲 67% [S5][S6]。階段一的 USD 400 也是同一套算法：28 至 32 支指的是留下來的測試片段，每支以片長上緣 15 秒、乘上同一個 3 至 5 次重試倍數計，等於一支背後實際生成 45 至 75 秒、約 USD 10.4 至 17.3；USD 400 可做 23 至 38 支，計畫取中段的 28 至 32 支，重試的錢已含在 400 裡面，額度用完階段一就停。
Further flags carried from the source workings: the claimed 30-second single-generation cap is unconfirmed on the API side (the budget does not depend on it) unconfirmed未確認; daily free points that could pad stage 1 are third-party reported, unverified on the official page unverified未核實; the API is in pre-release, with third-party resale prices ranging USD 0.09 to 0.134 per second pre-release, prices vary預發布期，轉售價不一.其餘從原始試算帶過來的標記：官方宣稱的單次 30 秒上限在 API 端未確認（本預算不依賴它）unconfirmed未確認；可補貼階段一的每日免費點數為第三方整理，未在官方頁核實 unverified未核實；API 仍在預發布期，第三方轉售價每秒 USD 0.09 至 0.134 不等 pre-release, prices vary預發布期，轉售價不一。
These two numbers sit together to give the budget a market coordinate, and that is all I use them for. I am deliberately not deriving a payback period or a savings figure from them, because I do not have defensible numbers for that kind of claim and refuse to present ones I cannot stand behind. What I can stand behind is the nature of the investment, in four points:這兩個數字放在一起，是為了給預算一個市場上的座標，我只把它們當座標用。我刻意不從這裡推導回本時間或省下多少錢這類結論，因為我手上沒有能支撐那種說法的數字，站不住的話我就選擇完全不寫。我站得住的是這筆投資的性質，四點：
Owned equipment keeps serving after the money is spent once; a subscription goes to zero the month it stops being paid.買斷的設備付一次錢就持續服務；訂閱制在停止付費的那個月就歸零。
Footage, style, and training data stay in the company's own hands, unaffected when a service provider changes rules, prices, or review standards.素材、風格與訓練資料留在公司自己手上，服務商改規則、改價、改審查標準都影響不到我們。
The hardware and everything it produces are company assets: the drives migrate to the company build, and every output lands on the company-data drive from day one.硬體與它產出的一切都是公司資產：硬碟會移交到公司機，每一件產出從第一天起就落在公司資料碟上。
Templates, style, and failure notes are compounding assets: the earlier they start accumulating, the thicker they get, and starting late means starting from zero.模板、風格與失敗紀錄是累積型資產：越早開始累積越厚，晚開始就等於從零再來。
During the transition the pipeline runs on my personal machine, and everything that migrates (drives, monitor, templates, workflows) moves to a company machine afterward, so the company keeps the assets it paid for. Separating personal and company data onto distinct drives is what protects both sides' privacy and security in the meantime. 過渡期產線先跑在我的個人電腦上，之後所有會跟著走的東西（硬碟、螢幕、模板、工作流程）整批移到公司電腦，公司花的錢買到的資產都會跟過去，不吃虧。過渡期間把個人與公司資料放在不同的實體碟上分流，就是同時保住雙方隱私與資安的做法。
Why the dedicated drive matters for security: company footage and credentials never mix with personal files, the whole company workload can be handed over or audited by moving one physical drive, and my personal data never has to leave with it. This section and the ladder's final rung in section 05 are the same plan: the drive bought at rung 2 is the vehicle the handover rides on.專用碟對資安的意義：公司素材與憑證從頭到尾不跟個人檔案混放，要移交或稽核時搬一顆實體碟就是完整交接，我的個人資料也不必跟著走。本節與第 05 節階梯的最後一階是同一個計畫：第 2 階買的那顆碟，就是整個移交所依託的載體。
Scope note, stated to avoid surprise later: the future company machine's spec, cost, and purchase date are outside this report. My suggestion is to propose that purchase after stage 2 proves the pipeline, at which point it will be a DDR5 build and the migrating items slot straight in.範圍先講清楚免得日後突襲：未來那台公司電腦的規格、費用與採購時間不在本報告範圍內。我的建議是等階段二驗證產線成立後再提那個採購案，屆時是 DDR5 平台，會跟著走的品項直接裝上去。
Everything below is flagged rather than silently assumed. Grouped by source.以下每一項都照實標記，沒有默默假設。依來源分組。
From the machine audit [S8]來自本機實測 [S8]
From the market survey (all quotes dated as marked, re-confirm before purchase) [S1][S2][S3][S4][S9]來自市場調查（查價日各如標示，採購前均需重新確認）[S1][S2][S3][S4][S9]
From the video-cost workings [S5][S6][S7]來自影片成本試算 [S5][S6][S7]
From the local-video verification [S11]來自本機影片生成查證 [S11]
Still in progress as this report ships本報告成稿時仍在進行的事