[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news:tech-deepseek-vs-openai":3},{"meta":4,"markdown":69,"quiz":70},{"type":5,"articleId":6,"slug":7,"title":8,"titleEn":9,"category":10,"summary":11,"publishedAt":12,"tags":13,"vocabulary":16,"quizId":68},"news","news-tech-deepseek-vs-openai","tech-deepseek-vs-openai","DeepSeek vs OpenAI 2026 — 中国AI企業の躍進と米国の対応","DeepSeek vs OpenAI in 2026 — The Rise of Chinese AI Firms and the US Response","tech","How the AI landscape shifted after DeepSeek's R1 release demonstrated that competitive frontier reasoning models could be trained at a fraction of the assumed cost. Covers the open-weights vs closed-weights split, the US semiconductor export-control response, and what Japanese enterprises should watch when choosing between OpenAI, DeepSeek, and domestic LLMs.\n","2026-04-20T00:00:00Z",[14,15],"ai","openai",[17,22,26,31,35,40,44,48,52,56,60,64],{"word":18,"reading":19,"meaning":20,"level":21},"躍進","やくしん","rapid advance","N1",{"word":23,"reading":24,"meaning":25,"level":21},"推論","すいろん","inference",{"word":27,"reading":28,"meaning":29,"level":30},"基盤","きばん","foundation","N2",{"word":32,"reading":33,"meaning":34,"level":30},"大規模","だいきぼ","large-scale",{"word":36,"reading":37,"meaning":38,"level":39},"学習","がくしゅう","training","N3",{"word":41,"reading":42,"meaning":43,"level":30},"公開","こうかい","open",{"word":45,"reading":46,"meaning":47,"level":30},"重み","おもみ","weight (of a neural network)",{"word":49,"reading":50,"meaning":51,"level":30},"性能","せいのう","performance",{"word":53,"reading":54,"meaning":55,"level":30},"効率","こうりつ","efficiency",{"word":57,"reading":58,"meaning":59,"level":39},"競争","きょうそう","competition",{"word":61,"reading":62,"meaning":63,"level":30},"採用","さいよう","adoption",{"word":65,"reading":66,"meaning":67,"level":21},"制約","せいやく","constraint","news-tech-deepseek-vs-openai-quiz","\n::para\n[中国]{ちゅうごく:Chinese:N5}の[AI]{えーあい:AI}[企業]{きぎょう:company:N1}[DeepSeek]{でぃーぷしーく:DeepSeek}が2025[年]{ねん:year:N5}[初頭]{しょとう:beginning of:N3}に[公開]{こうかい:released:N4}した[推論]{すいろん:reasoning:N1}[モデル]{もでる:model}「[R1]{あーるわん:R1}」は、[性能]{せいのう:performance:N3}と[コスト]{こすと:cost}の[両面]{りょうめん:both aspects:N3}で[業界]{ぎょうかい:industry:N4}に[衝撃]{しょうげき:shock:N1}を[与えた]{あたえた:gave:N3}と[報じられて]{ほうじられて:reported:N3}います。[本記事]{ほんきじ:this article:N3}では、その[後]{あと:aftermath:N5}の[流れ]{ながれ:flow:N3}と2026[年]{ねん:year:N5}[現在]{げんざい:currently:N3}の[競争]{きょうそう:competitive:N2}[構図]{こうず:landscape:N3}を[整理]{せいり:organize:N1}します。\n\n#en\nThe reasoning model \"R1,\" released by Chinese AI company DeepSeek at the beginning of 2025, is reported to have shocked the industry on both performance and cost dimensions. This article organizes the flow of events that followed and the competitive landscape as of 2026.\n::\n\n::heading\n[DeepSeek]{でぃーぷしーく:DeepSeek}[R1]{あーるわん:R1}[公開]{こうかい:release:N4}の[衝撃]{しょうげき:shock:N1}\n\n#en\nThe Shock of the DeepSeek R1 Release\n::\n\n::para\n[DeepSeek]{でぃーぷしーく:DeepSeek}[R1]{あーるわん:R1}は[OpenAI]{おーぷんえーあい:OpenAI}の[o1]{おーわん:o1}に[匹敵]{ひってき:comparable to:N1}する[推論]{すいろん:reasoning:N1}[性能]{せいのう:performance:N3}を、[公表]{こうひょう:reportedly:N3}される[学習]{がくしゅう:training:N4}[費用]{ひよう:cost:N3}が[従来]{じゅうらい:conventional:N1}の[フロンティア]{ふろんてぃあ:frontier}[モデル]{もでる:model}より[一桁]{ひとけた:order of magnitude:N1}[低い]{ひくい:lower:N2}[水準]{すいじゅん:level:N2}で[達成]{たっせい:achieved:N3}したと[されて]{されて:said to be}います。さらに[モデル]{もでる:model}の[重み]{おもみ:weights:N4}を[公開]{こうかい:openly released:N4}した[点]{てん:point:N3}が[特徴]{とくちょう:characteristic:N1}で、[研究者]{けんきゅうしゃ:researcher:N4}や[企業]{きぎょう:company:N1}が[自社]{じしゃ:own company:N4}[環境]{かんきょう:environment:N1}で[実行]{じっこう:run:N3}できる[大規模]{だいきぼ:large-scale:N1}[モデル]{もでる:model}の[選択肢]{せんたくし:option:N1}が[一気]{いっき:all at once:N5}に[広がった]{ひろがった:expanded:N4}と[評価]{ひょうか:evaluated:N1}されています。\n\n#en\nDeepSeek R1 is said to have achieved reasoning performance comparable to OpenAI's o1 at a reported training cost an order of magnitude lower than conventional frontier models. A further characteristic is that the model weights were openly released — and this is evaluated as having all at once expanded the options for large-scale models that researchers and companies can run in their own environments.\n::\n\n::callout\n[株式]{かぶしき:stock:N1}[市場]{しじょう:market:N3}では[米国]{べいこく:US:N3}の[半導体]{はんどうたい:semiconductor:N2}[関連]{かんれん:related:N3}[銘柄]{めいがら:stocks:N1}が[一時]{いちじ:temporarily:N5}[大幅]{おおはば:significantly:N2}に[下落]{げらく:fell:N3}するなど、[投資家]{とうしか:investors:N3}にも[強い]{つよい:strong:N4}[心理]{しんり:psychological:N4}[的]{てき:emotional:N4}[影響]{えいきょう:impact:N1}を[与えた]{あたえた:gave:N3}と[報じられて]{ほうじられて:reported:N3}います。\n\n#en\nIn the stock market, US semiconductor-related stocks reportedly fell significantly for a time, and a strong psychological impact on investors was reported.\n::\n\n::heading\n[オープン]{おーぷん:open}[ウェイト]{うぇいと:weight}と[クローズド]{くろーずど:closed}[ウェイト]{うぇいと:weight}の[二極化]{にきょくか:bifurcation:N2}\n\n#en\nThe Bifurcation Between Open-Weights and Closed-Weights\n::\n\n::para\n[業界]{ぎょうかい:industry:N4}は[OpenAI]{おーぷんえーあい:OpenAI}や[Anthropic]{あんすろぴっく:Anthropic}に[代表]{だいひょう:represented:N3}される[クローズド]{くろーずど:closed}[ウェイト]{うぇいと:weight}[陣営]{じんえい:camp:N1}と、[DeepSeek]{でぃーぷしーく:DeepSeek}や[Meta]{めた:Meta}の[Llama]{らま:Llama}に[代表]{だいひょう:represented:N3}される[オープン]{おーぷん:open}[ウェイト]{うぇいと:weight}[陣営]{じんえい:camp:N1}に[大きく]{おおきく:broadly:N5}[分かれ]{わかれ:divided:N5}つつあると[されて]{されて:said to be}います。[クローズド]{くろーずど:closed}[陣営]{じんえい:camp:N1}は[最先端]{さいせんたん:cutting-edge:N1}[モデル]{もでる:model}の[性能]{せいのう:performance:N3}と[安全]{あんぜん:safety:N3}[管理]{かんり:management:N2}で[優位性]{ゆういせい:advantage:N3}を[訴える]{うったえる:claim:N1}[一方]{いっぽう:on the other hand:N4}、[オープン]{おーぷん:open}[陣営]{じんえい:camp:N1}は[コスト]{こすと:cost}と[カスタマイズ]{かすたまいず:customization}[性]{せい:potential:N3}、[データ]{でーた:data}[主権]{しゅけん:sovereignty:N3}を[強み]{つよみ:strength:N4}として[企業]{きぎょう:enterprise:N1}[採用]{さいよう:adoption:N2}を[拡大]{かくだい:expand:N1}していると[見られて]{みられて:are seen:N5}います。\n\n#en\nThe industry is said to be increasingly bifurcating into a closed-weights camp represented by OpenAI and Anthropic and an open-weights camp represented by DeepSeek and Meta's Llama. The closed camp claims advantages in cutting-edge model performance and safety management, while the open camp is seen as expanding enterprise adoption with strengths in cost, customizability, and data sovereignty.\n::\n\n::heading\n[米国]{べいこく:US:N3}の[輸出]{ゆしゅつ:export:N2}[管理]{かんり:control:N2}[強化]{きょうか:strengthening:N3}\n\n#en\nThe US Strengthening of Export Controls\n::\n\n::para\n[米国]{べいこく:US:N3}は[DeepSeek]{でぃーぷしーく:DeepSeek}の[躍進]{やくしん:rapid advance:N1}を[受け]{うけ:in response to:N3}、[先端]{せんたん:advanced:N1}[GPU]{じーぴーゆー:GPU}や[製造]{せいぞう:manufacturing:N1}[装置]{そうち:equipment:N2}の[中国]{ちゅうごく:China:N5}[向け]{むけ:to:N3}[輸出]{ゆしゅつ:export:N2}[管理]{かんり:control:N2}を[更に]{さらに:further:N3}[強化]{きょうか:strengthen:N3}する[方針]{ほうしん:direction:N2}を[示して]{しめして:indicated:N3}いると[報じられて]{ほうじられて:reported:N3}います。[クラウド]{くらうど:cloud}[経由]{けいゆ:via:N3}での[GPU]{じーぴーゆー:GPU}[計算]{けいさん:compute:N2}[資源]{しげん:resource:N1}[利用]{りよう:use:N3}も[規制]{きせい:regulation:N3}の[対象]{たいしょう:target:N2}として[検討]{けんとう:considered:N1}されており、[中国]{ちゅうごく:China:N5}[企業]{きぎょう:company:N1}が[最先端]{さいせんたん:cutting-edge:N1}[モデル]{もでる:model}を[学習]{がくしゅう:train:N4}するのに[必要]{ひつよう:needed:N3}な[計算]{けいさん:compute:N2}[基盤]{きばん:infrastructure:N1}に[制約]{せいやく:constraints:N3}を[掛けよう]{かけよう:try to impose:N3}とする[動き]{うごき:moves:N4}が[続いて]{つづいて:continuing:N3}いると[見られて]{みられて:are seen:N5}います。\n\n#en\nIn response to DeepSeek's rapid advance, the US is reported to have indicated a direction of further strengthening export controls on advanced GPUs and manufacturing equipment to China. Use of GPU compute resources via the cloud is also reportedly being considered as a target of regulation, and moves to impose constraints on the compute infrastructure that Chinese companies need to train cutting-edge models are seen as continuing.\n::\n\n::heading\n[OpenAI]{おーぷんえーあい:OpenAI}と[Anthropic]{あんすろぴっく:Anthropic}の[対応]{たいおう:response:N1}\n\n#en\nThe Response from OpenAI and Anthropic\n::\n\n::para\n[OpenAI]{おーぷんえーあい:OpenAI}は[DeepSeek]{でぃーぷしーく:DeepSeek}の[公開]{こうかい:release:N4}を[受け]{うけ:in response to:N3}、[一部]{いちぶ:some:N3}[モデル]{もでる:model}の[利用]{りよう:use:N3}[料金]{りょうきん:pricing:N4}を[引き下げる]{ひきさげる:lowering:N3}とともに、より[小型]{こがた:smaller:N2}で[高速]{こうそく:high-speed:N3}な[モデル]{もでる:model}や[エージェント]{えーじぇんと:agent}[機能]{きのう:functionality:N3}の[強化]{きょうか:strengthening:N3}に[力]{ちから:effort:N4}を[入れて]{いれて:putting in:N5}いると[報じられて]{ほうじられて:reported:N3}います。[Anthropic]{あんすろぴっく:Anthropic}も[Claude]{くろーど:Claude}の[長文]{ちょうぶん:long-form:N4}[推論]{すいろん:reasoning:N1}や[コーディング]{こーでぃんぐ:coding}[支援]{しえん:support:N1}[性能]{せいのう:performance:N3}の[向上]{こうじょう:improvement:N3}を[継続]{けいぞく:continuing:N1}しているとされ、[モデル]{もでる:model}[層]{そう:tier:N2}ごとに[使い分け]{つかいわけ:differentiated use:N4}を[促す]{うながす:promote:N1}[戦略]{せんりゃく:strategy:N2}が[広がって]{ひろがって:spreading:N4}いると[見られて]{みられて:are seen:N5}います。\n\n#en\nIn response to DeepSeek's release, OpenAI is reported to have lowered pricing on some models, while putting effort into smaller, faster models and strengthening agent functionality. Anthropic is also said to be continuing to improve Claude's long-form reasoning and coding support performance, and a strategy of promoting differentiated use across model tiers is seen as spreading.\n::\n\n::heading\n[日本]{にほん:Japanese:N5}[企業]{きぎょう:enterprise:N1}の[選択]{せんたく:choice:N1}と[国産]{こくさん:domestic:N3}[LLM]{えるえるえむ:LLM}\n\n#en\nThe Choices Facing Japanese Enterprises and Domestic LLMs\n::\n\n::para\n[日本]{にほん:Japanese:N5}[企業]{きぎょう:company:N1}にとっては、[OpenAI]{おーぷんえーあい:OpenAI}や[Anthropic]{あんすろぴっく:Anthropic}の[API]{えーぴーあい:API}を[利用]{りよう:use:N3}するか、[DeepSeek]{でぃーぷしーく:DeepSeek}や[Llama]{らま:Llama}など[オープン]{おーぷん:open}[ウェイト]{うぇいと:weight}[モデル]{もでる:model}を[自社]{じしゃ:own company:N4}[環境]{かんきょう:environment:N1}に[置いて]{おいて:host:N3}[利用]{りよう:use:N3}するか、あるいは[ELYZA]{えらいざ:ELYZA}や[NEC]{えぬいーしー:NEC}の[cotomi]{ことみ:cotomi}など[国産]{こくさん:domestic:N3}[LLM]{えるえるえむ:LLM}を[採用]{さいよう:adopt:N2}するか、という[選択肢]{せんたくし:options:N1}が[広がって]{ひろがって:expanding:N4}いると[されて]{されて:said to be}います。[官公庁]{かんこうちょう:government agency:N2}や[金融]{きんゆう:financial:N1}[機関]{きかん:institution:N3}では、[データ]{でーた:data}[主権]{しゅけん:sovereignty:N3}や[セキュリティ]{せきゅりてぃ:security}[要件]{ようけん:requirements:N3}から[国産]{こくさん:domestic:N3}[モデル]{もでる:model}の[評価]{ひょうか:evaluation:N1}が[進んで]{すすんで:progressing:N3}いると[報じられて]{ほうじられて:reported:N3}います。\n\n#en\nFor Japanese firms, the options are reportedly expanding — using OpenAI or Anthropic APIs, hosting open-weights models such as DeepSeek or Llama in their own environment, or adopting domestic LLMs such as ELYZA or NEC's cotomi. In government agencies and financial institutions, evaluation of domestic models is reportedly progressing due to data sovereignty and security requirements.\n::\n\n::heading\n[今後]{こんご:going forward:N5}の[論点]{ろんてん:points of discussion:N3}\n\n#en\nPoints of Discussion Going Forward\n::\n\n::para\n[今後]{こんご:going forward:N5}は[モデル]{もでる:model}の[性能]{せいのう:performance:N3}そのものに[加え]{くわえ:in addition to:N3}、[推論]{すいろん:inference:N1}[コスト]{こすと:cost}、[エージェント]{えーじぇんと:agent}[能力]{のうりょく:capability:N3}、[安全]{あんぜん:safety:N3}[性]{せい:characteristic:N3}、[輸出]{ゆしゅつ:export:N2}[管理]{かんり:control:N2}の[対象]{たいしょう:target:N2}になるかどうかなど、[複数]{ふくすう:multiple:N2}の[軸]{じく:axes:N1}で[評価]{ひょうか:evaluated:N1}される[時代]{じだい:era:N4}に[入った]{はいった:entered:N5}と[見られて]{みられて:seen:N5}います。[米中]{べいちゅう:US-China:N3}[技術]{ぎじゅつ:technology:N2}[競争]{きょうそう:competition:N2}が[激化]{げきか:intensifies:N1}する[中]{なか:amidst:N5}、[日本]{にほん:Japanese:N5}[企業]{きぎょう:company:N1}は[特定]{とくてい:specific:N3}[ベンダー]{べんだー:vendor}[依存]{いぞん:dependence:N2}を[避け]{さけ:avoid:N1}つつ、[業務]{ぎょうむ:business:N3}[要件]{ようけん:requirements:N3}に[応じて]{おうじて:according to:N1}[最適]{さいてき:optimal:N3}な[組み合わせ]{くみあわせ:combination:N3}を[選ぶ]{えらぶ:choose:N3}[姿勢]{しせい:stance:N1}が[求められて]{もとめられて:required:N3}いると[されて]{されて:said to be}います。\n\n#en\nGoing forward, the era is seen as having entered one in which models are evaluated on multiple axes — model performance itself plus inference cost, agent capability, safety, and whether they fall under export controls. Amid intensifying US-China technology competition, Japanese firms are said to be required to take a stance of avoiding dependence on a specific vendor while choosing the optimal combination according to business requirements.\n::\n",{"id":68,"title":71,"titleEn":72,"topicPath":5,"questions":73},"DeepSeek vs OpenAI 確認テスト","DeepSeek vs OpenAI Confirmation Test",[74,101,123,146,169],{"id":75,"articleId":6,"question":76,"options":79,"correctLabel":85,"explanation":96,"tags":99},"news-tech-deepseek-vs-openai-quiz-q01",{"en":77,"jp":78},"What does the article cite as DeepSeek R1's biggest characteristic?","DeepSeek R1の最大の特徴として記事が挙げているものはどれか。",[80,84,88,92],{"label":81,"jp":82,"en":83},"ア","GPT-4より大きいパラメータ数","More parameters than GPT-4",{"label":85,"jp":86,"en":87},"イ","推論性能と低コストを両立し重みを公開した","Combined reasoning performance with low cost and open-released the weights",{"label":89,"jp":90,"en":91},"ウ","日本語専用に学習された","Was trained exclusively for Japanese",{"label":93,"jp":94,"en":95},"エ","OpenAIと共同で開発された","Was developed jointly with OpenAI",{"en":97,"jp":98},"The article cites three features: comparable reasoning performance, an order-of-magnitude lower training cost, and openly released weights.","記事は推論性能と桁違いに低い学習コスト、そして重み公開という三点を特徴として挙げている。",[100],"comprehension",{"id":102,"articleId":6,"question":103,"options":106,"correctLabel":85,"explanation":119,"tags":122},"news-tech-deepseek-vs-openai-quiz-q02",{"en":104,"jp":105},"Which pair represents the 'closed-weights camp'?","「クローズドウェイト陣営」に該当する企業の組み合わせはどれか。",[107,110,113,116],{"label":81,"jp":108,"en":109},"DeepSeek と Meta","DeepSeek and Meta",{"label":85,"jp":111,"en":112},"OpenAI と Anthropic","OpenAI and Anthropic",{"label":89,"jp":114,"en":115},"ELYZA と NEC","ELYZA and NEC",{"label":93,"jp":117,"en":118},"Meta と ELYZA","Meta and ELYZA",{"en":120,"jp":121},"The article cites OpenAI and Anthropic as representatives of the closed-weights camp, and DeepSeek and Meta (Llama) as the open-weights camp.","記事ではOpenAIとAnthropicがクローズドウェイト陣営、DeepSeekとMeta（Llama）がオープンウェイト陣営の代表として挙げられている。",[100],{"id":124,"articleId":6,"question":125,"options":128,"correctLabel":85,"explanation":141,"tags":144},"news-tech-deepseek-vs-openai-quiz-q03",{"en":126,"jp":127},"What measure did the US reportedly consider in response to DeepSeek's rise?","米国がDeepSeekの躍進を受けて検討していると報じられた措置はどれか。",[129,132,135,138],{"label":81,"jp":130,"en":131},"DeepSeekの米国上場の支援","Support for DeepSeek's US listing",{"label":85,"jp":133,"en":134},"中国向け先端GPUの輸出管理の更なる強化","Further strengthening of advanced GPU export controls to China",{"label":89,"jp":136,"en":137},"OpenAIの国有化","Nationalization of OpenAI",{"label":93,"jp":139,"en":140},"全AIモデルの公開義務化","Mandating that all AI models be open-sourced",{"en":142,"jp":143},"The article reports both strengthening GPU export controls and consideration of regulating GPU compute use via the cloud.","記事はGPU輸出管理の強化と、クラウド経由の計算資源利用も規制対象として検討されていると報じている。",[100,145],"policy",{"id":147,"articleId":6,"question":148,"options":151,"correctLabel":85,"explanation":164,"tags":167},"news-tech-deepseek-vs-openai-quiz-q04",{"en":149,"jp":150},"What does 'omomi' (重み) mean in context?","「重み」（おもみ）とは文脈上どういう意味か。",[152,155,158,161],{"label":81,"jp":153,"en":154},"サーバーの物理的な重量","Physical weight of the server",{"label":85,"jp":156,"en":157},"AIモデルのパラメータ","AI model parameters",{"label":89,"jp":159,"en":160},"ライセンス料金","License fee",{"label":93,"jp":162,"en":163},"重要度のランキング","Importance ranking",{"en":165,"jp":166},"In AI context, 'weights' refers to the trained parameters of a neural network. Releasing them publicly is what makes a model 'open-weight.'","AI文脈での「重み（weights）」はニューラルネットワークの学習済みパラメータを指す。これを公開するのが「オープンウェイト」モデル。",[168],"vocabulary",{"id":170,"articleId":6,"question":171,"options":174,"correctLabel":85,"explanation":187,"tags":190},"news-tech-deepseek-vs-openai-quiz-q05",{"en":172,"jp":173},"Why is evaluation of domestic LLMs progressing in Japanese government agencies and financial institutions?","日本企業がLLMを採用する際、官公庁や金融機関で国産モデル評価が進む主な理由は何か。",[175,178,181,184],{"label":81,"jp":176,"en":177},"国産モデルの方が性能が高いから","Domestic models perform better",{"label":85,"jp":179,"en":180},"データ主権とセキュリティ要件のため","Due to data sovereignty and security requirements",{"label":89,"jp":182,"en":183},"OpenAIが日本でサービスを停止したから","OpenAI stopped serving Japan",{"label":93,"jp":185,"en":186},"国産モデルの方が安いから","Domestic models are cheaper",{"en":188,"jp":189},"The article explicitly cites data sovereignty and security requirements — not performance or price.","記事はデータ主権とセキュリティ要件を理由として明記している。性能や価格ではない。",[100,191],"enterprise"]