crystal liu
)
輝達 Nvidia 清倉 AI 製藥
【當 NVIDIA 賣掉 Recursion，AI 製藥真正的考卷才剛開始】
Recursion Pharmaceuticals 在 2026 年 2 月 18 日上演了一場很有代表性的股價震盪。NVIDIA 最新 13F 揭露，截至 2025 年 12 月 31 日，它已經清空原本持有的 771 萬股 Recursion；消息一出，Recursion 盤中一度重挫約 14%，但當天尾盤買盤回來，股價最終反而收漲約 2%，隔天再續彈超過 5%。幾乎同一時間，Cathie Wood 旗下的 ARK 又再加碼約 125 萬股，而 ARK 最新揭露持股已來到約 3,732 萬股。這一退一進，並不是誰對誰錯那麼簡單，而是整個市場對 AI 製藥的估值邏輯，已經正式進入分歧期。
【NVIDIA 賣股，不等於判 Recursion 死刑】
很多人第一反應是：連 NVIDIA 都走了，Recursion 是否已經被判死刑？
但這種解讀過於粗糙。因為 NVIDIA 賣掉的是股權，不是雙方的技術合作。Recursion 在 2025 年第四季法說會上明確表示，與 NVIDIA 的技術合作仍在持續，重點仍放在自動化與超級運算；更何況，NVIDIA 在同一份 13F 中也同步退出了 Applied Digital、WeRide 等部位，顯示這更像投資組合調整，而不是針對 Recursion 單一公司的營運否定。
📌 真正值得注意的，不是「NVIDIA 有沒有賣」，而是市場現在開始把 AI 製藥從算力敘事，拉回臨床驗證與資本效率的現實考題。
【Recursion 為何會成為 TechBio 的代表樣本】
Recursion 之所以會成為這個賽道最具代表性的樣本，正因為它幾乎濃縮了過去十年 AI 製藥所有最迷人的想像。公司 2013 年成立，核心邏輯是把高通量細胞影像、機器視覺與深度學習結合，從表型出發理解生物學，再往藥物發現推進。
2023 年，NVIDIA 直接投資 5,000 萬美元，雙方宣布合作開發 AI 藥物發現基礎模型；2024 年 5 月，Recursion 再推出由 63 套 DGX H100 系統驅動的 BioHive-2，號稱是全球製藥產業自有且自營速度最快的超級電腦之一，運算速度較前一代快四倍。
從那一刻開始，Recursion 不再只是生技公司，更被視為「TechBio 樣板」。
【它不只講平台，還想把 biology 跟 chemistry 全部接起來】
接著它又做了另一個更大的動作：在 2024 年 8 月宣布以全股票方式收購 Exscientia，交易價值約 6.88 億美元，並在同年 11 月完成合併。
這筆交易的戰略含義很清楚：Recursion 想補足自己在生物學與表型資料上的優勢，接上 Exscientia 在化學設計、分子優化與自動化合成端的能力，打造一個從 biology 到 chemistry、從資料到分子的垂直整合平台。
這種併購也反映出 AI 製藥公司的商業模式正在快速混血：不再只是賣工具，不再只是做合作，也不再只講自研管線，而是三者同時進行。
【但真正的考題，從來都不是 GPU，而是臨床】
問題是，理想的整合平台，最後還是得面對生技產業最古老也最殘酷的現實：臨床不是靠算力就能跳過的關卡，現金流也不是靠故事就能無限支撐。
Recursion 在 2025 年 5 月大幅重整管線，停掉 REC-2282、REC-994 與 REC-3964 等計畫，並在 6 月宣布裁員約 20%。這背後不是單一事件，而是公司承認：即便擁有大模型、超級電腦與豪華合作陣容，資本市場最終仍只會為高機率穿越臨床死亡谷的資產買單。
這一點，Recursion 自己的財報其實已經寫得很明白。公司 2025 年全年營收為 7,470 萬美元，主要仍來自合作收入；同期淨虧損則擴大到 6.448 億美元。雖然截至 2025 年底帳上現金、現金等價物與受限現金合計約 7.539 億美元，管理層也表示在不額外融資下，現金跑道可延伸到 2028 年初，但這些數字同時說明了一件事：Recursion 至今仍不是一間靠產品銷售自我造血的藥企，而是一間正在用合作收入、資本市場耐心與現金儲備，爭取第一個真正可驗證臨床成功樣本的 TechBio 公司。
【市場耐心為什麼開始變少】
而市場耐心之所以被快速消耗，原因也不難理解。
Recursion 先前最受矚目的 REC-994，是用於 symptomatic cerebral cavernous malformation（CCM，症狀性腦海綿狀血管畸形）的候選藥物。它在 2024 年公布的中期臨床數據中，安全性與耐受性達標，MRI 也出現部分病灶縮小訊號，但患者與醫師可感知的臨床改善並不夠明確，療效路徑始終沒有真正被釐清；到了 2025 年，這項計畫最終被停掉。
另一個重要項目 REC-2282，原本用於 NF2-mutated meningioma（NF2 突變腦膜瘤），也因整體腫瘤縮小與臨床活性不足而被放棄。
⚠️ 這兩件事都提醒投資人：AI 可以更快找到候選分子、縮短前端搜尋時間，但它無法自動把模糊的臨床生物學，變成清楚的病人獲益。
【REC-4881，為什麼會被市場重新放大】
也正因如此，REC-4881 才會被市場重新放大。
這個原名 TAK-733 的分子，是 Recursion 2020 年自 Takeda 引進的 MEK1/2 變構抑制劑，後來透過自家平台重新定位到 familial adenomatous polyposis（FAP，家族性腺瘤性息肉症）。
從公司 2025 年底公布的 TUPELO 試驗資料來看，REC-4881 在 12 週治療後，75% 可評估患者的總息肉負荷下降，中位數下降 43%；更關鍵的是，在停藥 12 週後，11 位可評估患者中仍有 82% 維持反應，中位數息肉負荷較基線下降 53%。
公司已規劃於 2026 年上半年與 FDA 啟動註冊路徑溝通，並預計於 2027 年上半年釋出更多 Phase 1b/2 資料。Recursion 甚至直接把這個案子定義為「Recursion OS 的第一個臨床驗證」。
✨ 這不是小事，因為它代表市場終於看到這個平台不只是能產生靶點與分子，還可能真的產生對病人有意義的臨床結果。
【AI 製藥真正的分水嶺，已經不是模型大小】
這也是 AI 製藥今天最值得被講清楚的真相：產業真正的分水嶺，不再是誰擁有更大的模型、更多的 GPU，或更會講基礎模型故事，而是誰能把平台能力轉成臨床可驗證資產。
從這個角度看，Recursion 並不孤單。Schrödinger 2025 年財報就很清楚展現了商業模式混血化：全年軟體收入 1.995 億美元，藥物發現收入也已達 5,640 萬美元，代表最成功的「工具型公司」也早就不只賣軟體。另一邊，Recursion 透過合作收入與自研管線並行，Exscientia 則直接與它合體。
今天的 AI 製藥，已經很難再用 SaaS、CRO、Biotech 三個盒子切乾淨。
真正把產業拉開差距的，反而是臨床結果。Insilico Medicine 在 2025 年把 Rentosertib（ISM001-055）於 idiopathic pulmonary fibrosis（IPF，特發性肺纖維化）的 Phase 2a 結果發表在《Nature Medicine》，並把它定位為 AI 驅動藥物發現的第一個臨床概念驗證之一；另一端，Verge Genomics 的 ALS 候選藥 VRG50635 則在預先設定的療效分析未過關後被終止，公司重新退回技術平台與服務模式。
這一正一反，幾乎就是整個產業的縮影：AI 的確能提高前端發現效率，但能不能穿越人體試驗，仍然是另一套完全不同的競爭。
【NVIDIA 這次賣股，真正提醒了市場什麼】
所以，NVIDIA 這次撤出 Recursion，真正暴露的不是「AI 製藥不行了」，而是這個產業終於不再能只靠一頁簡報、一套算力敘事或一個明星股東活著。
市場現在要看的，是更硬的三件事：
🔹 第一，你能不能反覆產生值得進臨床的資產。
🔹 第二，你的第一個成功能不能帶出第二個、第三個，而不是只是一發煙火。
🔹 第三，在臨床結果還沒完全兌現之前，你有沒有足夠的資本紀律活到那一天。
NVIDIA 賣掉 Recursion 股權，不代表它否定了 AI 藥物發現；它只是提醒所有人，GPU 可以加速搜尋，卻不能替代機轉、終點、病人反應與財務現實。
【寫在最後：AI 製藥的估值邏輯，正在成熟】
Recursion 接下來究竟會成為 AI 製藥的先烈，還是第一批真正穿越死亡谷的倖存者，現在還沒有定論。
但有一件事已經很明確：AI 製藥的估值邏輯，正在從「會不會改變世界」轉向「能不能做出真正的藥」。
而這，才是整個產業開始成熟的標誌。
參考資料：
[0]:公開資料＆各公司官網
[1]: https://investor.nvidia.com/.../sec.../default.aspx... "Financial Info - SEC Filings"
[2]: https://www.fool.com/.../recursion-rxrx-q4-2025-earnings.../
[3]: https://www.reuters.com/.../nvidia-invests-50-mln.../
[4]: https://www.reuters.com/.../biotech-firm-recursion-buy.../
[5]: https://www.sec.gov/.../000160183026000039/rxrx-20251231.htm "rxrx-20251231"
[6]: https://www.globenewswire.com/.../recursion-reports... "Recursion Reports Fourth Quarter and Full Year 2025"
[7]: https://www.reuters.com/.../recursions-blood-vessel.../
[8]: https://www.sec.gov/.../000160183026000039/rxrx-20251231.htm
[9]: https://ir.schrodinger.com/.../Schrdinger.../default.aspx "Schrödinger, Inc. - Schrödinger Reports Fourth Quarter and Full-Year 2025 Financial Results"
[10]: https://insilico.com/.../tnrecuxsc1-insilico-announces... "Insilico Announces Nature Medicine Publication of Phase IIa Results of Rentosertib, the Novel TNIK Inhibitor for Idiopathic Pulmonary Fibrosis Discovered and Designed with a Pioneering AI Approach"
Trent Nvidia Clearance AI Pharmaceutical
【When NVIDIA sells Recursion, the real test for AI medicine is just starting 】
Recursion Pharmaceuticals staged a representative stock price shock on February 18, 2026. NVIDIA's latest 13F revealed that as of December 31, 2025, it had emptied 7.71 million shares of Recursion; as soon as the news came out, Recursion fell about 14 percent during the day, but the stock ended up rising about 2 percent, and the next day renewed more than 5 percent. Almost at the same time, Cathie Wood's ARK added another 1.25 million shares, and ARK's latest disclosed holding has reached about 37.32 million shares. This step back is not that simple as who is right and who is wrong, but the valuation logic of the entire market for AI medicines has officially entered a period of disagreement.
【Nvidia's stock sale does not mean Recursion's death sentence 】
Many people's first reaction is: Even NVIDIA is gone, has Recursion been sentenced to death?
But this interpretation is too rough. Because NVIDIA sold equity, not technical cooperation. Recursion made it clear at the fourth quarter of 2025, that the technical cooperation with NVIDIA is continuing, with the focus still on automation and supercomputing; not to mention that NVIDIA also withdrew from Applied Digital, WeRide, etc. in the same 13F, showing that this is more like a portfolio adjustment than a negative for Recursion’s single company.
📌 What’s really worth noting is not “Is Nvidia selling it” but the market is now starting to pull AI medicine from the narrative of computing power back to the realistic questions of clinical verification and capital efficiency.
【Why Recursion is a representative sample of TechBio 】
Recursion is the most iconic sample of this track precisely because it condensates almost all of the most fascinating imaginations of AI medicine over the past decade. Founded in 2013, the core logic of the company is to combine high-throughput cell imaging, machine vision and deep learning to understand biology from the phenotype and advance to drug discovery.
In 2023, NVIDIA directly invested $50 million, and the two parties announced a partnership to develop AI drug discovery basic models; in May 2024, Recursion launched BioHive-2, powered by 63 DGX H100 systems, claiming to be one of the fastest supercomputers in the global pharmaceutical industry, running four times faster than the previous generation.
From that moment on, Recursion is no longer just a biotech company, but a “TechBio model”.
【It's not just about platform, but also wants to connect biology and chemistry 】
Then it made another bigger move: it announced an all-stock acquisition of Exscientia in August 2024 worth approximately $688 million, and completed the merger in November of the same year.
The strategic implications of this deal are clear: Recursion wants to complement its strengths in biology and phenotype data, connect with Exscientia’s capabilities in chemical design, molecular optimization, and automated synthesis, and build a platform for vertical integration from biology to chemistry, data to molecules.
This acquisition also reflects the rapid mix of AI pharmaceutical companies’ business models: no longer just selling tools, no longer just collaborating, no longer just talking about self-development pipeline, but all three at the same time.
[But the real test question is never GPUs, it's clinical]
The problem is that the ideal integration platform will end up facing the oldest and most cruel reality of the biotech industry: clinical is not a level that can be skipped by computing power, and cash flow is not an infinite support by stories.
Recursion significantly reorganised its pipeline in May 2025, halting REC-2282, REC-994 and REC-3964 programs, and announced a layoff of approximately 20% in June. Behind this is not a single event, but an admission by the company that even with big models, supercomputers, and luxury collaborations, capital markets will end up paying only for high-probability assets crossing clinical death valley.
Recursion's own financial report is actually very clear on this. The company's full year 2025 revenue of $74.7 million, still largely from partnership revenue; net loss expanded to $644.8 million over the same period. While cash, cash equivalent and restricted cash together totaled about $753.9 million at the end of 2025, and management said the cash runway could be extended to early 2028, with no additional financing, these figures also say one thing: Recursion is not a drug company that sells its own hematopoietic products, but a TechBio company that is using partnership income, capital market patience, and cash reserves to secure the first truly verifiable clinical success sample.
【Market patience why starts to get less 】
The reason why market patience is being consumed quickly is not hard to understand.
Recursion, the most high-profile REC-994, is a drug candidate for symptomomatic cerebral cavernous malformation (CCM). Its mid-term clinical data released in 2024, safety and tolerability were met, and MRI showed some signs of lesion shrinkage, but the clinical improvements perceived by patients and doctors were not clear enough, and the path of efficacy was never really cleared; by 2025, the program was finally discontinued.
Another important project, REC-2282, originally for NF2-mutated meningioma (NF2 mutated meningioma), was also abandoned due to overall tumor shrinkage and inadequate clinical activity.
⚠ Both of these remind investors that AI can find candidate molecules faster and reduce front-end search times, but it can’t automatically turn blurry clinical biology into clear patient benefits.
【REC-4881, why will be re-enlarified by the market 】
This is why REC-4881 is re-enlarified by the market.
This molecule, originally named TAK-733, is a MEK1/2 allosteric inhibitor introduced by Recursion from Takeda in 2020, and later repositioned to familial adenomatous polyposis (FAP, familial adenomatous polyposis).
According to the TUPELO trial data released by the company at the end of 2025, REC-4881 reduced total polyp load in 75% of assessable patients after 12 weeks of treatment, and the median decreased 43%; more importantly, 82% of the 11 assessable patients still maintained a response after 12 weeks of discontinuation, and the median polyp load decreased 53% from the baseline.
The company has planned to communicate with the FDA for the initiation of registration path in the first half of 2026, and expects to release more Phase 1b/2 data in the first half of 2027. Recursion even defined the case directly as "the first clinical verification of Recursion OS".
✨ This is no small thing, because it represents the market is finally seeing that this platform not only produces targets and molecules, but may actually produce clinical results that are meaningful to patients.
【AI Pharmaceuticals real hydrate, no longer model size 】
This is also the truth that AI Pharmaceuticals is most worth telling today: The real divide for the industry is no longer who has bigger models, more GPUs, or more basic model stories, but who can turn platform capabilities into clinically verifiable assets.
From this perspective, Recursion is not alone. Schr ödinger's 2025 financial report clearly shows the mixed business model: $199.5 million in software revenue for the year, $56.4 million in drug discovery revenue, representing the most successful "tool companies" that are not just selling software. On the other side, Recursion runs side by side with the self-development pipeline through collaboration revenue, Exscientia fits directly with it.
Today's AI medicine, it's hard to cut clean with three boxes of SaaS, CRO, Biotech.
What really pulls the industry apart is the clinical result. Insilico Medicine published the Phase 2a results of Rentosertib (ISM001-055) on idiopathic pulmonary fibrosis (IPF) in Nature Medicine in 2025, positioning it as one of the first clinical proof-of-concepts for AI-driven drug discovery; on the other end, Verge Genomics' ALS candidate VRG50635 was terminated after pre-set efficacy analysis was passed, and the company returned to the technology platform and service model.
This is almost the epitome of the entire industry: AI can indeed improve the efficiency of front-end discovery, but whether it can pass through human trials is still a completely different set of competition.
【Nvidia's stock sale really reminds the market of what 】
So, NVIDIA's withdrawal from Recursion this time is not that "AI medicine is not dead", but that the industry is finally no longer able to live on just one page of presentations, a set of computing narratives, or a star shareholder.
What the market is looking at now are three harder things:
🔹 First, can you repeatedly generate assets worth getting into the clinic.
🔹 Second, can your first success lead to a second or third, not just a fireworks.
🔹 Third, do you have enough capital discipline to live until the clinical results are fully realized?
NVIDIA's sale of Recursion doesn't mean it negates AI drug discovery; it just reminds everyone that GPUs can speed up search, but not a substitute for mechanism, endpoint, patient response, and financial reality.
【Writed at the end: The valuation logic of AI medicine is maturing 】
Whether Recursion will be the pioneer of AI medicine next or the first survivors to actually cross Death Valley is still uncertain.
But one thing is clear: the valuation logic of AI medicine is moving from “will it change the world” to “can it make real drugs?”
And this is the sign that the whole industry is beginning to mature.
Reference material:
[0]:Public information & official website of each company
[1]: https://investor.nvidia.com/.../sec.../default.aspx... "Financial Info - SEC Filings"
[2]: https://www.fool.com/.../recursion-rxrx-q4-2025-earnings.../
[3]: https://www.reuters.com/.../nvidia-invests-50-mln.../
[4]: https://www.reuters.com/.../biotech-firm-recursion-buy.../
[5]: https://www.sec.gov/.../000160183026000039/rxrx-20251231.htm "rxrx-20251231"
[6]: https://www.globenewswire.com/.../recursion-reports... "Recursion Reports Fourth Quarter and Full Year 2025"
[7]: https://www.reuters.com/.../recursions-blood-vessel.../
[8]: https://www.sec.gov/.../000160183026000039/rxrx-20251231.htm
[9]: https://ir.schrodinger.com/.../Schrdinger.../default.aspx "Schrödinger, Inc. - Schrödinger Reports Fourth Quarter and Full-Year 2025 Financial Results"
[10]: https://insilico.com/.../tnrecuxsc1-insilico-announces... "Insilico Announces Nature Medicine Publication of Phase IIa Results of Rentosertib, the Novel TNIK Inhibitor for Idiopathic Pulmonary Fibrosis Discovered and Designed with a Pioneering AI Approach"