[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-eu-ai-gigafactory-7-factories-100000-chips":3,"news-related-ae924988-53ef-4f46-b1c2-c43c4e9866fe":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"ae924988-53ef-4f46-b1c2-c43c4e9866fe","欧盟掏 100 亿欧元建 7 座 AI 超级工厂：每座堆 10 万颗顶尖芯片，瞄准美国算力代差","欧盟委员会启动 7 座 AI 超级工厂计划，公开 100 亿欧元公共资金招标，希望撬动 200 亿欧元私人投资。每座工厂至少配 10 万颗尖端 AI 芯片，性能约为当前欧盟数据中心 4 倍。这是欧盟与美国算力差距对标的关键一步。","# 欧盟掏 100 亿欧元建 7 座 AI 超级工厂：每座堆 10 万颗顶尖芯片，瞄准美国算力代差\n\n## 一笔 100 亿欧元的\"算力赌注\"\n\n欧盟 27 国集体下场，把 AI 基础设施摆到了国家级战略的位置。当地时间周四，欧盟委员会正式开放 7 座 AI 超级工厂（AI Gigafactory）的招标通道：100 亿欧元公共资金打底，希望撬动 200 亿欧元私人投资，单座工厂要求至少部署 10 万颗尖端 AI 芯片，性能大约是目前欧盟运行数据中心平均水平的 4 倍。\n\n这是一个明确的\"对标动作\"。按美国 2025 年对欧洲 AI 能力的评估，欧洲在前沿模型、算力规模、芯片供给三方面全面落后于美国；今年 6 月提交给欧洲议会的一份委员会内部报告也直言，欧洲企业和公共部门会持续依赖美国 AI 提供商，这反过来又压住了本土前沿服务商的成长空间。\n\n## 为什么是\"超级工厂\"而不是数据中心\n\n超级工厂的设计目标不是跑普通的云服务，而是把训练下一代前沿大模型所需的三件事——算力、电力、数据——打包到同一物理空间里：\n\n- **算力密度**：每座 10 万颗尖端 AI 芯片的规模直接对标 xAI Colossus、微软 \u002F OpenAI Stargate 这类美方旗舰集群；\n- **能效与电力**：超级工厂选址会绑定电网容量和绿电指标，避免出现\"芯片到位电跟不上\"的尴尬；\n- **主权数据**：与此前欧盟已落地的 13 座 \"AI Factory\"（轻量级超算 + 数据资源）相比，这次的 \"Gigafactory\" 明确要承接对主权数据敏感的模型训练任务，让欧洲企业不必为了合规把数据传到美国云上。\n\n这意味着 100 亿欧元并不是补贴，而是\"入场费\"——欧盟把入场券发给愿意共担长期投资的企业，再用私人资本把杠杆撑到 300 亿欧元体量。\n\n## 谁会来竞标\n\n虽然正式投标通道刚开，潜在玩家已经能画出一张清单：\n\n- **本土超算 \u002F 云服务商**：EuroHPC 联盟成员、Atos \u002F OVHcloud 这类已经运营 EuroHPC 机器的厂商是天然候选；\n- **欧洲 AI 头部**：Mistral AI（法国）、Aleph Alpha（德国）、Stability AI 的欧洲业务线都需要本土算力，对主权云需求极强；\n- **美国巨头欧洲子公司**：微软、AWS、Google 在欧洲有大规模数据中心扩建计划，但欧盟很可能在合规上对它们额外审查；\n- **芯片供给方**：英伟达、AMD 是最大赢家，但欧盟同步在推 European Processor Initiative 与 RISC-V 路线，长期会要求一定比例的本土芯片采购。\n\n## 钱从哪儿来、卡在哪里\n\n100 亿欧元来自欧盟多年财政预算框架（MFF）的数字欧洲与创新基金切片，加上复苏与韧性基金（RRF）的剩余额度。说实话，这笔钱对单座 10 万卡集群来说并不算多——业内估算一座前沿训练集群的硬件 + 电力 + 冷却总投入在 30 亿到 50 亿欧元之间，100 亿欧元摊到 7 座头上，每座只有 14 亿左右，杠杆几乎全部押在私人投资那一块。\n\n更现实的卡点有三个：\n\n1. **电力审批**：一座 10 万卡集群满载运行通常需要数百兆瓦电力，欧洲电网扩容周期 3–5 年，发电指标的获取比芯片更难；\n2. **芯片采购**：H100 \u002F B100 级别的高端 GPU 实际交付周期长达 12–18 个月，加上出口管制变数，欧盟必须在采购清单上写明\"不依赖单一供应商\"；\n3. **人才虹吸**：DeepMind 这周刚被曝 AlphaFold 核心团队三人跳槽 Anthropic，说明欧洲在 AI 研究人才上输给美国前沿实验室的趋势还在延续；没有顶尖研究者，再多集群也会空转。\n\n## 真正的\"所以呢\"\n\n超级工厂计划不是单纯的基础设施投资，它本质上是欧盟给本土 AI 产业的一份\"安全垫\"：\n\n- 对 Mistral 这种欧洲头部来说，意味着终于有可能不靠 Azure \u002F AWS 就跑出下一代前沿模型；\n- 对监管者来说，本土算力是 EU AI Act 落地的前提——你不能一边要求数据不出境，一边把训练算力全押在境外云上；\n- 对英伟达 \u002F AMD 来说，欧盟把\"撬动私人资本\"写进招标条款，相当于让芯片厂商以更优条件进入一个长期采购清单；\n- 对中国 AI 厂商来说，这是欧洲一次明确的\"减少对美依赖\"信号，但在出口管制大背景下，欧盟不太可能让中国厂商直接接入这套基础设施。\n\n接下来 6–12 个月值得盯三件事：哪些财团中标、首批工厂选址是否落在北欧 \u002F 南欧绿电富集区、以及欧盟会不会在第二批招标里加入\"必须训练一个欧洲主权基础模型\"的硬性 KPI。如果超级工厂只是变成了又一个\"政府补贴的数据中心升级包\"，那 100 亿欧元大概率会打水漂；如果它真能孵化出对标 GPT \u002F Claude 级别的欧洲前沿模型，这一轮才算真正改变了 AI 力量格局。","https:\u002F\u002Fwap.cj.sina.cn\u002Fpc\u002F7x24\u002F5016629","cfc3f7ed-9911-4f50-9532-1255a5b2d178",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ff00120c-9efd-4b2c-a2ad-e1f173ab73fa","en","EU commits EUR 10B to 7 AI gigafactories to close the compute gap","The European Commission has opened bidding for 7 AI gigafactories backed by €10 billion in public funds, aiming to leverage an additional €20 billion in private investment. Each site will host at least 100,000 cutting-edge AI chips and roughly 4x the performance of typical EU data centers today. The plan is the blocs most explicit response to widening transatlantic compute inequality.","# EU Puts Up €10B for 7 AI Gigafactories: 100,000 Top-Tier Chips Per Site, Aimed at Closing the Compute Gap With the US\n\n## A €10 billion compute gamble\n\nThe EUs 27 member states have elevated AI infrastructure to a national-strategy level. On Thursday, the European Commission formally opened bidding for 7 AI Gigafactories: €10 billion in public funds as anchor capital, with the goal of mobilizing another €20 billion in private investment. Each site is required to host at least 100,000 cutting-edge AI chips, and the aggregate compute capacity should land at roughly 4x the average EU data center in operation today.\n\nThis is an explicit benchmarking move. Per the US 2025 assessment of European AI capabilities, Europe trails the US across frontier models, compute scale, and chip supply. An internal Commission report submitted to the European Parliament in June was even more blunt: European enterprises and public-sector bodies will continue to depend on US AI providers, which in turn suffocates the growth of European frontier service providers.\n\n## Why \"Gigafactory\" rather than data center\n\nThe gigafactory is not designed to run ordinary cloud workloads. It packages the three things training the next generation of frontier models requires — compute, power, and data — into a single physical footprint:\n\n- **Compute density**: 100,000 top-tier chips per site puts the EU in the same conversation as xAI Colossus and the Microsoft \u002F OpenAI Stargate flagship clusters.\n- **Energy and grid**: Site selection is tied to grid capacity and green-energy availability, to avoid the recurring \"chips arrive before the power does\" failure mode.\n- **Sovereign data**: Unlike the 13 lighter-weight \"AI Factories\" the EU has already deployed (modest supercomputers plus curated data resources), the Gigafactories are explicitly intended to host sovereignty-sensitive model training, so European enterprises do not have to ship regulated data to US hyperscalers to access frontier capability.\n\nThat reframes the €10 billion: it is not a subsidy, it is an entry ticket. The Commission hands the ticket to consortia willing to co-invest long term, and private capital stretches the leverage to roughly €30 billion.\n\n## Who is likely to bid\n\nThe bidding window has just opened, but the shortlist of credible bidders is already visible:\n\n- **Domestic HPC and cloud providers**: EuroHPC consortium members, Atos, OVHcloud — operators already running EuroHPC machines are the natural anchor tenants.\n- **European AI leaders**: Mistral AI (France), Aleph Alpha (Germany), the European business lines of Stability AI all need sovereign compute and have an unusually strong demand pull for it.\n- **US hyperscaler European subsidiaries**: Microsoft, AWS, and Google all have large European data center build-outs in flight, but expect the Commission to apply extra compliance scrutiny.\n- **Chip suppliers**: Nvidia and AMD are the biggest near-term winners, but the Commission is simultaneously funding the European Processor Initiative and RISC-V tracks; a domestic-sourcing quota is likely in the longer-term contracts.\n\n## Where the money comes from, and where it gets stuck\n\nThe €10 billion is drawn from the Digital Europe Programme and parts of the Innovation Fund inside the EUs Multiannual Financial Framework, plus residual capacity from the Recovery and Resilience Facility. Honestly, that figure is not large for what is being asked: industry estimates put a single frontier training cluster at €3–5 billion once hardware, power, and cooling are included. Spread across 7 sites, the public envelope leaves roughly €1.4 billion per site, which means almost all of the leverage sits on the private-capital side.\n\nThree practical bottlenecks will decide whether this actually works:\n\n1. **Power and permitting**: A fully-loaded 100,000-GPU cluster typically draws several hundred megawatts. European grid expansion cycles run 3–5 years, and generation permits are harder to secure than chip allocations.\n2. **Chip procurement**: H100 \u002F B100-class accelerators carry 12–18 month lead times even before export-control complications. The procurement clauses will need to explicitly forbid single-vendor dependence.\n3. **Talent drain**: This very week DeepMind saw three AlphaFold core members leave for Anthropic, which underlines that Europe is still losing AI research talent to US frontier labs. Even a perfect data center goes to waste without top-tier researchers to run it.\n\n## The real \"so what\"\n\nThe Gigafactory program is not a plain infrastructure spend. It is, at its core, a safety net the EU is laying under its domestic AI industry:\n\n- For Mistral and similar European leaders, it finally makes it possible to train the next frontier model without depending on Azure or AWS.\n- For regulators, domestic compute is a precondition for the EU AI Act to function — you cannot simultaneously require data residency and push all training compute into foreign clouds.\n- For Nvidia and AMD, the Commissions \"leverage private capital\" clause effectively offers chip vendors preferential access to a long-term procurement pipeline.\n- For Chinese AI vendors, this is a clear European signal of intent to reduce US dependence — but under current export-control architecture, the EU is unlikely to allow Chinese firms to plug directly into this infrastructure.\n\nThree things to watch over the next 6–12 months: which consortia win the bids, whether the first wave of sites lands in the green-power-rich Nordic or Southern European regions, and whether the Commission adds a hard KPI in the second tranche requiring the training of a sovereign European foundation model. If the gigafactories turn out to be just another subsidized data center upgrade, the €10 billion is largely wasted. If they really do incubate a European frontier model at GPT \u002F Claude tier, this round genuinely redraws the AI power map.","eu-ai-gigafactory-7-factories-100000-chips","2026-07-31T04:30:00Z","2026-07-30T20:04:06.350727Z","2026-07-30T20:04:06.350737Z",true,"agent",215,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"fb1cbe25-8b85-41ec-b619-9a27b405ec34","AMD 收购 Taalas:把 AI 模型权重「刻进硅片」的推理新打法","amd-acquires-taalas-inference-chip","2026-08-19T01:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"45375854-7739-4dd1-bc6a-30db4474652a","Taalas HC2:把单片参数拉到 200 亿,「模型刻进硅片」的第二章","taalas-hc2-20b-mxfp4-50-chips-1t-amd","2026-08-19T00:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"dfdc3216-52aa-4a78-9bf5-859affc37d17","AMD 收下 Taalas：把模型权重刻进芯片，推理的内存墙还剩多少？","amd-acquires-taalas-msic-etched-weights","2026-08-11T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"c07c67b6-6a48-4780-88bd-bc46b628c546","AMD 吃下 Taalas:把模型权重永久刻进芯片的\"硬推理\"赌局","amd-taalas-hardwired-inference-aug-2026","2026-08-08T12:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"2434bbc6-4fda-4750-a02d-dd3ca1fe8933","AMD 收购 Taalas:把模型权重刻进芯片,押注推理硬件的\"硬核\"路线","amd-acquires-taalas-hardcore-inference-silicon","2026-08-08T04:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"bb1d01e1-c61f-4428-b91a-41000a26bec5","从「堆硬件」到「卖Token」:10余家上市公司押注Token工厂,算力行业 TaaS 模式浮出水面","china-ai-token-factory-taas-shift-2026","2026-07-31T04:00:00+00:00"]