[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-xai-colossus-2gw-grok-4-6-7-compute":3,"news-related-70b5b0d6-ce28-48e8-abe4-6a667a723c4e":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},"70b5b0d6-ce28-48e8-abe4-6a667a723c4e","xAI 把 Colossus 推到 2 GW:555,000 颗 GPU 撑起 Grok 4.6\u002F4.7 的万亿参数竞速","马斯克把 Colossus 数据中心扩建到 2 GW 总功率,555,000 颗 NVIDIA GPU 落户孟菲斯,代号 MACROHARDRR 的第三栋楼正式入列。这套堆栈正是 Grok 4.6(1.5T)与 4.7(2.1T)两个月内连发的训练底座,xAI 用规模换迭代速度,把算力即护城河的逻辑推到极致。","## 技术背景:从 Colossus 1 到 Colossus 2 的 19 天奇迹\n\n2024 年,xAI 在田纳西州孟菲斯把一座旧工业厂房改造成 AI 训练基地,这就是 Colossus 的起点。当时 NVIDIA CEO 黄仁勋用「超人」形容其交付速度——从选址到首批 GPU 上线,只用 19 天,而传统数据中心建设周期动辄 2-3 年。秘诀不是赶工,而是「绕开电网」:xAI 在园区里自建燃气发电,避开 ERCOT 式的并网排队,把电力、冷却、网络三层基础设施全部垂直整合。\n\nColossus 1 已经容纳 230,000 颗 GPU(其中 30,000 颗 GB200),耗电约 500 MW。Colossus 2 在此基础上直接翻倍:55 万颗 GB200\u002FGB300,总功率冲到 1 GW。两者相加,孟菲斯园区已经是全球最大的单点 AI 训练设施。\n\n## 核心事件:MACROHARDRR 把总功率顶到 2 GW\n\n2025 年 12 月 30 日,马斯克在 X 上确认 xAI 买下密西西比州 Southaven 的第三栋建筑,代号 **MACROHARDRR**(延续他对微软的「Macrohard」命名梗)。这栋楼紧邻 Colossus 2,允许 xAI 用超低延迟网络把三处设施连成一个统一计算环境。\n\n新园区数据如下:\n\n- **总 GPU 规模**:约 555,000 颗,采购成本约 180 亿美元(平均每颗约 32,400 美元)\n- **总功率**:~2 GW,等效于 150 万户美国家庭的用电量\n- **GPU 构成**:GB200 约 52 万颗 + GB300 约 3 万颗 + 早期 H100\u002FH200 约 3 万颗\n- **冷却**:液冷强制散热,需要每分钟 50,000+ 加仑冷却水(密西西比河流域供水)\n\n横向对比一下,这个体量是 Meta AI 研究中心(~500 MW)的 4 倍,Microsoft Azure AI(~400 MW)的 5 倍。Musk 公开说过 xAI 的目标是「拥有比所有人都多的 AI 算力」——现在他在单点上做到了。\n\n## Grok 4.6 \u002F 4.7:用算力换迭代速度\n\n这套堆栈不是摆设,而是直接服务于 xAI 的模型节奏。Musk 7 月 18 日在 X 上确认 Grok 4.6 进入发布管线,7 月 24 日进一步给出硬时间表:4.6 两周内、4.7 再两周后,两个万亿参数模型在 4 周内连发。\n\n具体规格:\n\n- **Grok 4.6**:1.5 万亿参数,V9 基座,主打 SFT + RL 升级,预计 8 月 7 日上线\n- **Grok 4.7**:2.1 万亿参数,Musk 自称「每个维度都比 4.6 强,只是推理稍慢」\n\n对比一下,GPT-3 是 1750 亿参数——xAI 现在单模型已经是它的 10 倍以上。而 Grok 4.5 已经在 SWE Marathon 编程基准上拿到 29.0%,领先 Claude Opus 4.8 的 26.0%。Colossus 2 提供的不只是「更大的模型」,而是「更频繁的训练-评估-发布」闭环:Musk 之前透露 xAI 维持每周两次的模型更新频率,2 GW 算力是支撑这种节奏的物理底座。\n\n## 行业影响:算力即护城河\n\nxAI 的玩法给整个前沿 AI 实验室划了一条新的基线:在 OpenAI、Anthropic、Google 都还在 1 GW 量级时,xAI 已经把单点算力顶到 2 GW。这不只是一个数字游戏——更快的训练循环意味着可以同时跑多个实验候选、可以承受 RL 微调更高的失败成本、可以在更长上下文 \u002F 更大 MoE 上做更激进的尝试。\n\n但代价同样显眼。Colossus 自建燃气电厂的模式在田纳西\u002F密西西比引发电网压力和环保争议,「AI 工厂」对当地水资源和电费的冲击已经不是隐忧。如果 4.6\u002F4.7 的发布真能兑现「更便宜、Opus 级」的承诺,xAI 的算力赌注就赢了;如果性能不及预期,180 亿美元的 GPU 库存就会变成全行业最贵的固定资产。\n\n## 所以呢\n\n短期看,Grok 4.6\u002F4.7 是 xAI 对「参数即能力」路线的最新验证——V9 基座 + 1.5T\u002F2.1T 参数 + 升级 RL,配合 2 GW 训练底座,这是 2025 年下半年最激进的 frontier 模型冲刺。长期看,这场算力军备竞赛正在重新定义「头部实验室」的入场券:没有 100 万颗 GPU 级别的训练设施,就别想在前沿模型榜上保持位置。Musk 已经把门槛抬到了 555,000 颗,下一步,看 OpenAI 和 Anthropic 怎么接。","https:\u002F\u002Fx.ai\u002Fcolossus","b82e17a3-1dbd-4b5d-88dc-9f518f917cc0",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"6467a397-ce8f-4de2-a3e4-b3d6bbc8cc33","en","xAI's Colossus hits 2GW: 555,000 GPUs for the trillion-param race","Musk expanded the Colossus data center complex to 2 GW total power, with 555,000 NVIDIA GPUs landing in Memphis and a third building—codenamed MACROHARDRR—joining the cluster. This stack is the training substrate behind Grok 4.6 (1.5T) and Grok 4.7 (2.1T), both arriving within two months. xAI trades raw scale for iteration speed, taking the 'compute is moat' thesis to its logical extreme.","## Background: From Colossus 1 to the 19-day Colossus 2 miracle\n\nIn 2024, xAI converted an old industrial building in Memphis, Tennessee into an AI training facility—the original Colossus. NVIDIA CEO Jensen Huang called its delivery 'superhuman': from site selection to the first GPUs coming online in just 19 days, versus the typical 2-3 year buildout cycle for traditional data centers. The trick wasn't overtime, but 'grid avoidance': xAI built on-site gas-fired generation, bypassing ERCOT-style interconnection queues and vertically integrating power, cooling, and networking.\n\nColossus 1 already houses 230,000 GPUs (including 30,000 GB200s) and draws ~500 MW. Colossus 2 doubled down: 550,000 GB200\u002FGB300 units, peak power hitting ~1 GW. Combined, the Memphis campus is already the largest single-site AI training facility on the planet.\n\n## Core event: MACROHARDRR pushes total capacity to 2 GW\n\nOn December 30, 2025, Musk confirmed on X that xAI bought a third building in Southaven, Mississippi—codenamed **MACROHARDRR** (continuing his 'Macrohard' naming jab at Microsoft). The site sits adjacent to Colossus 2, letting xAI wire the three facilities into a single, unified compute fabric via ultra-low-latency networking.\n\nThe new campus numbers:\n\n- **Total GPU count**: ~555,000, purchased for ~8B (averaging ~2,400 per GPU)\n- **Total power**: ~2 GW, equivalent to powering ~1.5 million U.S. households\n- **GPU mix**: ~520,000 GB200 + ~30,000 GB300 + ~30,000 legacy H100\u002FH200\n- **Cooling**: Liquid cooling mandatory at this density, requiring 50,000+ gallons of water per minute (sourced from the Mississippi River watershed)\n\nFor context, this is 4x Meta's AI Research Center (~500 MW) and 5x Microsoft Azure AI (~400 MW). Musk publicly stated xAI's goal is to 'have more AI compute than everyone else'—and now, on a single site, he's done it.\n\n## Grok 4.6 \u002F 4.7: Trading compute for iteration speed\n\nThis stack isn't decorative—it's the training substrate for xAI's model cadence. Musk confirmed Grok 4.6 was in the pipeline on July 18, then tightened the schedule on July 24: 4.6 in two weeks, 4.7 two weeks after that. Two trillion-parameter models shipping within a four-week window.\n\nConcrete specs:\n\n- **Grok 4.6**: 1.5 trillion parameters, V9 base, with significantly upgraded SFT + RL. Target release: around August 7.\n- **Grok 4.7**: 2.1 trillion parameters, which Musk describes as 'better than 4.6 in every way, except slightly slower to serve.'\n\nTo put the scale in perspective: GPT-3 was 175 billion parameters. xAI's single models are now more than 10x larger. Grok 4.5 already scores 29.0% on the SWE Marathon coding benchmark—beating Claude Opus 4.8's 26.0%. Colossus 2 delivers not just 'bigger models' but a tighter train-evaluate-ship loop: Musk has previously confirmed xAI maintains a twice-weekly update cadence. The 2 GW stack is what physically makes that rhythm possible.\n\n## Industry impact: Compute is the moat\n\nxAI's playbook sets a new baseline for frontier AI labs: while OpenAI, Anthropic, and Google are still in the ~1 GW range, xAI has pushed single-site compute to 2 GW. This isn't just a numbers game—faster training cycles mean you can run more experimental candidates in parallel, absorb higher RL fine-tuning failure costs, and make more aggressive bets on longer contexts and bigger MoE configurations.\n\nBut the costs are visible. Colossus's on-site gas generation model is already straining local grids and drawing environmental scrutiny in Tennessee and Mississippi; the impact of 'AI factories' on regional water and electricity prices is no longer hypothetical. If 4.6\u002F4.7 deliver on the 'Opus-level-but-cheaper' promise, xAI's compute gamble pays off. If they underdeliver, 8B of GPU inventory becomes the most expensive fixed asset in the industry.\n\n## So what\n\nIn the short term, Grok 4.6\u002F4.7 is xAI's latest validation of the 'parameters = capability' doctrine—V9 base, 1.5T\u002F2.1T parameters, upgraded RL, all backed by a 2 GW training substrate. It's the most aggressive frontier-model sprint of late 2025. In the long term, this compute arms race is redefining the entry ticket for 'frontier lab' status: without a million-GPU-class training facility, you don't get to stay on the leaderboard. Musk has set the bar at 555,000. Next move: OpenAI and Anthropic.","xai-colossus-2gw-grok-4-6-7-compute","2026-07-31T04:00:00Z","2026-07-30T22:04:55.440052Z","2026-07-30T22:04:55.440068Z",true,"agent",281,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"f4af1e4e-98a3-4810-831c-699ffe31ae73","马斯克公布 Grok 4.6\u002F4.7 路线图：1.5T\u002F2.1T 参数，SFT+RL 升级，8 月 7 日发行","grok-4-6-4-7-roadmap-1-5t-2-1t","2026-07-30T08:45:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"1d80585e-c797-4aa1-ac68-ef87334d5d0c","PLaMo 3.0 Prime 正式发布：PFN 把「日语实战」做成日本国产 LLM 的差异化战场","plamo-3-0-prime-pfn-japanese-domestic","2026-06-24T08:15:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"dbff301b-4dda-4537-8c3f-19ee4a6fd88e","字节跳动正训练 10 万亿参数模型:规模上已与 Anthropic Mythos 5 相当","bytedance-10t-parameter-model-pretraining","2026-08-11T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"8bd5a96a-b85b-4db6-ad54-a2c311867178","字节跳动被曝训练10万亿参数超大模型：对标Anthropic Mythos,中国LLM进入\"10T俱乐部\"前夜","bytedance-10-trillion-parameter-model","2026-08-07T09:11:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"ad10985b-425c-4af1-9495-c63792a2b593","腾讯混元把语音识别打到 3% WER：Hy ASR 3.0 preview 让 ASR 从“逐字”走向“读语境”","tencent-hunyuan-hy-asr-3-0-preview-context-aware","2026-08-05T00:00:00+00:00"]