[新聞] OpenAI 首款自研晶片預計交由台積電生產

[DrowningPool (My broken dreams)]

原文標題：Exclusive-OpenAI set to finalize first custom chip design this year

原文連結：https://reurl.cc/M6xa8L

發布時間：2025-2-10

記者署名：Anna Tong, Max A. Cherney and Krystal Hu

原文內容：

OpenAI is pushing ahead on its plan to reduce its reliance on Nvidia for its
chip supply by developing its first generation of in-house
artificial-intelligence silicon.

The ChatGPT maker is finalizing the design for its first in-house chip in the
next few months and plans to send it for fabrication at Taiwan Semiconductor
Manufacturing Co, sources told Reuters. The process of sending a first design
through a chip factory is called "taping out."

OpenAI and TSMC declined to comment.

The update shows that OpenAI is on track to meet its ambitious goal of mass
production at TSMC in 2026. A typical tape-out costs tens of millions of
dollars and will take roughly six months to produce a finished chip, unless
OpenAI pays substantially more for expedited manufacturing. There is no
guarantee the silicon will function on the first tape out and a failure would
require the company to diagnose the problem and repeat the tape-out step.

Inside OpenAI, the training-focused chip is viewed as a strategic tool to
strengthen OpenAI's negotiating leverage with other chip suppliers, the
sources said. After the initial chip, OpenAI's engineers plan to develop
increasingly advanced processors with broader capabilities with each new
iteration.

If the initial tape out goes smoothly, it would enable the ChatGPT maker to
mass-produce its first in-house AI chip and potentially test an alternative
to Nvidia's chips later this year. OpenAI’s plan to send its design to TSMC
this year demonstrates the startup has made speedy progress on its first
design, a process that can take other chip designers years longer.

Big tech companies such as Microsoft and Meta have struggled to produce
satisfactory chips despite years of effort. The recent market rout triggered
by Chinese AI startup DeepSeek has also raised questions about whether fewer
chips will be needed in developing powerful models in the future.

The chip is being designed by OpenAI’s in-house team led by Richard Ho,
which had doubled in the past months to 40 people, in collaboration with
Broadcom. Ho joined OpenAI more than a year ago from Alphabet's Google where
he helped lead the search giant's custom AI chip program. Reuters first
reported OpenAI’s plans with Broadcom last year.

Ho's team is smaller than the large-scale efforts at tech giants such as
Google or Amazon. A new chip design for an ambitious, large-scale program
could cost $500 million for a single version of a chip, according to industry
sources with knowledge of chip design budgets. Those costs could double to
build the necessary software and peripherals around it.

Generative AI model makers like OpenAI, Google and Meta have demonstrated
that ever-larger numbers of chips strung together in data centers make models
smarter, and as a result, they have an insatiable demand for the chips.

Meta has said it will spend $60 billion on AI infrastructure in the next year
and Microsoft has said it will spend $80 billion in 2025. Currently, Nvidia’
s chips are the most popular and hold a market share of roughly 80%. OpenAI
is itself participating in the $500 billion Stargate infrastructure program
announced by U.S. President Donald Trump last month.

But rising costs and dependence on a single supplier have led major customers
such as Microsoft, Meta and now OpenAI to explore in-house or external
alternatives to Nvidia’s chips.

OpenAI's in-house AI chip, while capable of both training and running AI
models, will initially be deployed on a limited scale, and primarily for
running AI models, the sources said. The chip will have a limited role within
the company's infrastructure.

To build out an effort as comprehensive as Google or Amazon's AI chip
program, OpenAI would have to hire hundreds of engineers.

TSMC is manufacturing OpenAI's AI chip using its advanced 3-nanometer process
technology. The chip features a commonly used systolic array architecture
with high-bandwidth memory (HBM) - also used by Nvidia for its chips - and
extensive networking capabilities, sources said.

OpenAI 正加速自研 AI 晶片，以降低對 NVIDIA 的依賴，預計今年內完成設計並送交台
積電（TSMC）生產，目標 2026 年量產。該晶片主要用於 AI 推論，採 3 奈米製程、高
頻寬記憶體（HBM）與 systolic array 架構，具備強大網路能力。開發團隊由前
Google AI 晶片專家 Richard Ho 領導，並與 Broadcom 合作，團隊已擴增至 40 人。開
發一款高階 AI 晶片成本可能達 5 億美元。由於 AI 晶片需求激增，Microsoft、Meta
及 OpenAI 等企業正尋求 NVIDIA 以外的解決方案。若首款晶片 tape-out 成功，
OpenAI 可能於今年內測試其作為 NVIDIA 替代方案。此舉不僅提升議價能力，也有助於
降低 AI 運算成本，強化 OpenAI 的技術自主權。

心得/評論：

OpenAI預計晶片交給台積電代工
新聞消息一出
夜盤直接嘎上天
世界的G怎麼輸？

空蛙表示：

--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 218.32.44.219 (臺灣)
※ 文章網址: https://www.ptt.cc/bbs/Stock/M.1739186057.A.A4C.html
→ baka1412    : 不然要給誰
推 tsubasawolfy: 不然你要給INTEL嗎?
推 Waitaha     : 真的只需要買2330就夠了
推 yayohola    : 不然還有誰能做
推 sonyvaio    : AVGO設計？
推 watashiD    : 要前往STARGATE了
→ lovepork    : deepseek:
推 greedypeople: 資本支出沒有調升的話這邊多別的地方就會少吧
推 phxcon      : 怎麼不早說?
→ Gipmydanger : https://i.imgur.com/7Ik3BkJ.jpg
→ afflic      : 其實不說也知道
推 julian42    : 輝達崩
推 onekoni     : https://i.imgur.com/m04vcsM.jpeg
推 jim543000   : 真的有多蛙被這種白癡東西騙進去接盤？
→ tsukuyomii  : AVGO未來趨勢大贏家
→ jameswang278: 世界的大GG，傳家寶
→ wr          : 阿不給台積電就剩三星給intl啦
→ bobsun      : 要收關稅了喔
→ ilGroundhog : 只要台積解決了關稅股價就水到渠成
推 truelove356 : AZ AZ AZ
→ cocoaswifty : 蛋糕沒有變大啊，是在嗨啥
推 mumuwei     : 師傅：沒什麼
推 kmshy       : 不然要給誰？
推 TAYLOR888   : 直接找工廠拿貨比較便宜的概念
推 julian42    : 蛋糕一樣大，但少了一個人在中間賺一手啊
推 ll6a        : 少了一個仲介，賣房的賺到，買房也賺到
推 truelove356 : 亞歷桑那亞歷桑那亞歷桑那
推 KIMBEOM     : 那就是創意受惠了
推 tigerzz3    : 剛台積博通一起噴
→ tigerzz3    : 還在創意 剛博通已經先動
推 qoosmsm     : 不然給誰做？記得後面乖乖排隊
推 QooSnow     : （請問價格）（貴或便宜？）（嘻嘻。）
推 amos30627   : 當初不是聽說要用A16 太貴了？
→ wr          : 每個巨頭一定都想過跳過nv找台積 但nv現在還是龍頭
推 BJ3906      : 創意明天鎖死
推 afacebook   : NV晶片有那麼好做 AMD早就做出來跟NV競爭了
推 a10181018   : 當NV AMD這麼多年的發展是吃素的？最後還是灰頭土
→ a10181018   : 臉回去找NV
→ ledmstony   : 創意要噴爛了嗎
推 hkcdc       : 不然還有誰
→ iusocute    : 不然還能給誰 沒台G等著被打趴嗎
推 junior020486: TSM噴啊
→ payneblue   : 每個巨頭都想搞自研，又不想設計外流，基本上代工只
→ payneblue   : 有一個選擇
推 jingbaby77  : 誘多 等我阿川講話就下去
推 onekoni     : 找三星等等被偷技術
→ stocktonty  : 台積電特點就是他們對客戶的東西保密做得很好
→ onekoni     : 不自己生產 只能找可信的雞
推 truelove356 : 跟創意？
→ ck326       : 只有一個選擇就是等著吃反壟斷套餐
推 ru04hj4     : AVGO+TSM
推 good10740   : NPU也能跑AI
推 tigerzz3    : 未來創意的份額不夠大 應該還是博通漫威吃掉
→ tigerzz3    : 最穩還是買AVGO
→ Feting      : ...不然能給誰，這樣能漲？
→ eternal0327 : 這麼快就能流片了嗎
推 dostey      : 10000%完全不會影響到
→ z23061542   : 台gg:單來我就接 好不好用是你設計問題
→ Kevinsa     : https://imgur.com/a/tsyoXK1
推 nickname1o1 : ODM ，散熱表示：
推 blueseason  : 今天南下載客，到底誰沒上車？
推 takuma      : 京元電 還有我啊
推 chobit618   : 參興：給個機會QQ
推 PureAnSimple: GG又又又贏了
推 ots625      : 今天誰被洗下車？
推 NiNiNiNiNi  : 創意要噴啦！
推 bigbill     : https://i.imgur.com/VAFl6Kq.jpeg
推 Eopay       : 一直講創意.. 裡面有提到創意嗎…
→ Eopay       : 上面那張圖只是大摩的預測 不準吧
推 DIDIMIN     : 靠背，什麼都給台積電做，等等川普又不開心
推 dragonjj    : 靠邀不然要給誰生產？你找博通脈維爾世芯創意都要
→ dragonjj    : 給台積電代工啊
推 DA3921999   : 誰今天沒加碼台積電的？
推 ImHoluCan   : 自研要是那麼簡單搞的好，巨頭人人可以自己搞顯卡
推 Toge        : 嗯，聽說川普要加稅了，嘻嘻
推 westmoney   : 明天又要漲了，天天數錢真是苦惱
推 mainsa      : 講AMD的應該沒搞懂自製跟外購的差別 我自製就是只
→ mainsa      : 要付台積電5成毛利 但外購要先付設計商5成毛利 他
→ mainsa      : 們再付GG5成毛利 等於一個晶片成本如果是100鎂 我
→ mainsa      : 自製的可以200買到 跟AMD買要400 跟NV買可能要500
→ mainsa      : 如果自製的效率不要輸NV太多就有搞頭(簡單講就是如
→ mainsa      : 果能做到AMD水平就可以考慮自製 但如果要往外買AMD
→ mainsa      : 水平是不夠的) 但現在先進製程的入門門檻超高 只有
→ mainsa      : 用量超大的巨頭公司有機會自己搞 其他規模小一點的
→ mainsa      : 還是只能乖乖買NV就是了
→ tsgd        : 現在就沒有其它家可以做 外星人放棄了INTEL 可惜啊
噓 jmcgow11    : 這根本廢話 不然誰還能做？
→ tsgd        : 自研是不容易啊 但你不投資 就要面對老黃漲價的臉色
推 jiansu      : 推論 有機會用低價的硬體 自己設計也不一定比量產
→ jiansu      : 划算 現在是nv利潤高 自己做可能比較便宜也避免被
→ jiansu      : 卡技術 改天說不定被搞到更低價 推論用成熟製程ed
→ jiansu      : ge就處理完了 外接pcie就處理好了
→ aimlikenoob : 賣鏟子的根本不擔心誰成為AI霸主
→ poco0960    : 大不了不放米國
→ poz93       : 訂單多 股東會宣布蓋美國二廠
推 Lowpapa     : 沒用了
推 skyhelp     : 自製只要給台積電五成毛利，那asic廠商免費？
推 TheTimes    : 問OPENAI，XAI 如何投資，小韮們 唯一選擇私募公募
→ TheTimes    : 兼得的Cross ETF - XOVR。
推 a64186401   : GG這麼保密 怎會流出這消息
推 wsx26997785 : 英特爾：美國同胞救救我
推 joggingwang : Open 要降低 NV需求，自研晶片，這2家都要找gg，除
→ joggingwang : 了一開始要試研期間對gg總投片數有增加，可是最後gg
→ joggingwang : 總量也沒增加多少吧
推 joggingwang : 這個新聞要是再以往發佈是利多，可是在現時期，變
→ joggingwang : 成又給三普勒索的藉口
推 hang520     : 不管哪一家都要找GG GG怎麼輸？
推 rodion      : 有些“工作者”真是說鬼話不嫌累喔 ＧＧ最擔心的
→ rodion      : 並非川普 而是太過單一的客戶分佈 有人來跟ＮＶ搶
→ rodion      : ＡＩ晶片市場 這對ＧＧ是再好不過的事情
→ rodion      : 因為 這就讓晶片製造的供需更往賣方市場推了
推 rodion      : 也就是說 不管川普政府要玩什麼花樣 ＧＧ的籌碼只會
→ rodion      : 因其於市場更全面的重要性提升而更多
推 shingatter  : 說好救救intel呢？
推 alphish     : 好喔~噴
→ WTF55665566 : 印tel就扶不起的阿斗 想救就要花大錢 馬監國跟川普
→ WTF55665566 : 怎麼可能多拿錢給它？
推 kducky      : 不然要給誰？
推 r40491101   : intel就是被太多印度人給害死的，跟MS一樣...
→ r40491101   : 大家都在用舌頭跳舞，沒人做事
推 truelove356 : 那些說創意鎖的 在哪
→ raku        : 不然要給誰 台灣代工晶片加關稅只是搞美國自己的fa
→ raku        : bless晶片公司
推 stosto      : 阿不就博通那顆3nm客制的asic
→ stosto      : 隨著客戶需求調整啊，直接下單博通不就好了，至少
→ stosto      : 博通搞那顆三四年了