步步惊心：丽

581亿，北大校友又干出一个AI独角兽_我的网站

一 |     AI算力“卖水人”的造富神话仍在继续。         近日，AI云平台Together AI宣布完成8亿美元C轮融资，投后估值飙升至83亿美元（合约581亿元）。

二 |     KYIV, Ukraine -- Russia appears to be suffering severe losses as it wages its new offensive in eastern Ukraine, amid some of the bloodiest fighting in months, according to Ukrainian officials and independent analysts.Russia has thrown thousands of troops and hundreds of tanks and armored vehicles into the offensive, which is focused mainly on the strategic city of Avdiivka and further north near the city of Kupiansk. Russian forces for roughly two weeks have been trying to encircle Avdiivka, which is one of the most heavily fortified areas of Ukraine's frontline.Ukrainian officials on Sunday claimed Russia had lost over 6,000 soldiers in a week, as well as over 400 armored vehicles and tanks. Britain's defense intelligence claimed Russian casualties have spiked by 90% since it began the offensive.ABC cannot independently verify those numbers, but numerous videos of the Russian attacks around Avdiivka released by Ukrainian units suggest Russia is taking losses on a scale not since last winter.Some of the videos show columns of Russian armor coming under withering Ukrainian fire as they try to advance, pummelled by artillery and anti-tank missiles, and running into minefields. In other videos, Russian forces can be seen being barraged by U.S.-supplied cluster munitions. Large groups of Russian infantry can be seen being struck by artillery.Open source analysts watching the videos have counted dozens of Russian vehicles destroyed and damaged.Ukraine does not publish its own casualties, but Ukrainian officials have acknowledged they are also large.So far Russian forces have made little progress, only advancing a few hundred meters, near Avdiivka, according to both Ukrainian and Russian accounts.But despite the losses, Russia is continuing its efforts to encircle Avdiivka. The Washington, D.C.-based Institute for the Study of War on Sunday assessed Russia was moving more troops up to continue the attack."Russian forces are funneling additional forces to the Avdiivka front despite ongoing challenges with frontal mechanized assaults and the failure of a renewed push on October 19-20," the ISW wrote in its daily update on the war.Ukrainian officials have said they believe the determined Russian effort to seize Avdiivka is motivated by the Kremlin's desire to claim a significant victory ahead of Russia's presidential elections next year."It appears that Russian forces are continuing to use ineffective and costly tactics in offensive efforts near Avdiivka leading to high personnel and vehicle losses, and that they have continuously done so in several waves of attacks over the past week suggests that the Russian military command is prioritizing this axis, despite continued and growing losses," the ISW wrote in an earlier assessment on Saturday.Some Russian pro-war military bloggers have criticised how the operation has been handled and the decision to attack the most heavily fortified area of Ukraine's line, given the small progress so far.Analysts, however, have warned that despite the catastrophic losses, Russia may still be able to encircle Avdiivka and that it still has sizeable reserves in the area.Vitaliy Barabash, the head of Avdiivka's Ukrainian military administration, on Monday said Russian forces were attempting to cut the only road leading to the city, which would isolate the Ukrainian forces defending it."Logistics are very difficult. Twenty two kilometers to the city are constantly under fire, both during the day and at night. It greatly complicates both evacuation and bringing in humanitarian aid. Any logistics is all through one road. Of course, the enemy is trying to cut it," Barabash said in a live interview with RFE/RL.He also said that 1,601 civilians still in Avdiivka did not want to leave.Ukraine is also still on the offensive. In the past week, an unusually large force from two elite Ukrainian Marine brigades landed on the Russian-occupied eastern bank of the Dneipr river in the Kherson region. The Ukrainian units have advanced into at least two villages there and are digging in, according to Russian military bloggers.The scale of the operation is still very unclear, but it appears Ukraine may be attempting to establish a bridgehead there that if successful could create a possibility for it to move larger forces across for a bigger attack, potentially opening a new front in its counteroffensive. For now, most analysts assess the size of the Ukrainian force operating on the eastern bank is still small.Ukrainian forces are also continuing to attack on the main axis of the counteroffensive in the southern Zaporizhzhia region and around the city of Bakhmut, reportedly making small advances.。本轮融资由沙特阿美旗下Aramco Ventures领投，英伟达（NVIDIA）、Vista Equity Partners、General Catalyst等明星机构跟投。         当前，在大模型竞争激烈之下，这家成立仅4年的AI云平台服务商，为何能吸引巨头们争相入局？答案，或许藏在“开源模型”与“华人技术天团”的两个关键词里。         一年估值上涨1.5倍，开源AI成吸金利器          Together AI成立于2022年，几乎与ChatGPT同期诞生。成立时，公司就明确定位——做一家 “AI原生云”公司 ，核心业务是向企业出租基于英伟达GPU集群的算力，让开发者能以极低成本运行开源大模型。         发展初期，以OpenAI为首的闭源模型，曾一度占据主导地位，但很快市场发现，企业端落地的AI业务面临一个残酷的事实：闭源前沿模型Token费用高昂，无情地吞噬了AI应用们的利润空间。而随着DeepSeek等高性能开源模型的强势崛起，行业现状正在发生改变。         在从“闭源”向“开源”转型的过程中，Together AI直击这一痛点，通过整合DeepSeek、MiniMax、Kimi等主流开源模型，为企业级客户提供Serverless推理、专属基础设施及批量推理等多元化服务，成为这波开源模型生态爆发中，最大的受益者。

三 |          依托完善的开源模型生态布局，Together AI打造出极具性价比的开源模型“平替”方案。数据显示，在同等甚至更优性能表现下，企业接入Together AI算力服务后，推理成本可降低6倍至60倍。         以其AI客服平台Decagon为例，当工作负载全部迁移至Together AI后，推理成本直接缩减6了倍。强大的成本优势，直接转化为真实的业绩订单：最新一季年度预订额显示，Together AI的订单量较上个季度，强势突破11.5亿美元，即从最开始的2024年初3000万美元，暴涨到现在的11.5亿美元，涨幅达38倍。         两年暴增38倍的年度预订额，彻底引爆了Together AI在开源AI领域的商业潜力。一是，随着AI Agent的爆发，企业AI已不再需要简单的问答机器人，而是能够处理复杂工作流的数字员工，这从中催生了海量的Token调用量。

四 | 二是，Together AI自身也通过底层技术的创新，打破算力成本的物理瓶颈。         其通过自主研发的自研ATLAS推理引擎，以及推测解码（Speculative Decoding）等前沿技术，将部分推理工作负载速度最高提升400%。这种全链路的技术优化，使得Together AI能够以远低于传统云厂商的成本，交付出企业级所需的推理服务，构筑起技术壁垒。         也正因如此，Together AI估值持续攀升。2025年2月，在由General Catalyst领投的B轮融资中，Together AI的估值仅33亿美元。

五 | 然而，16个月之后，其投后估值就达到83亿美元，这相当于一年内上涨了1.5倍。         北大校友与斯坦福教授坐镇，一支“华人技术天团”浮现          Together AI的快速崛起，除了踩准开源AI的产业风口，背后还有一支“华人核心技术团队”浮现。         翻看Together AI的官网，不难发现，其创始人团队堪称“学霸天团”，且华人占比较高。         首先作为这支技术天团的核心代表之一，联合创始人兼CTO的张策（Ce Zhang）来自于北京大学。履历显示，他拥有扎实的顶尖学术背景，2008年张策本科毕业于北大数学系，随后在威斯康星大学麦迪逊分校获得博士学位，且担任过苏黎世联邦理工学院计算机科学系的助理教授。         在研究中，他始终围绕如何让机器学习更便宜、更可信、更容易被更多人使用的这一主线，认为AI不应当只是巨头们的专属工具，而是通过系统层、调度层、分布式架构的创新，把训练和推理的算力成本打下来，让个人、初创公司、传统行业都能低成本地使用开源大模型。         基于这一核心理念，张策主导了Together AI在底层算力调度与推理引擎上的创新，带领团队研发自研的ATLAS推理引擎，并深度集成了NVIDIA Dynamo等先进的分布式推理技术。这一系列的创新，让Together AI实现了上下文感知的智能路由与KV缓存管理，有效避免了大规模算力部署中的冗余计算问题，进而在商业上成功斩获年度预订额突破11.5亿美元的亮眼成绩。         其次，张策之外，公司联合创始人珀西·梁也是一名深耕AI领域的顶尖华人学者。身为斯坦福大学计算机科学系的教授，珀西·梁是斯坦福基础模型研究中心（CRFM）的主任。在自然语言处理与机器学习领域，他专注于开源基础模型的研发、模型评测体系的搭建等。         在团队合作上，如果说，张策在底层架构方面，为Together AI搭建了一条高效、低成本的算力“高速公路”，那么珀西·梁则凭借深厚的学术积累与行业认知，成为公司技术落地的“导航仪”。他通过严谨的评测基准，帮助Together AI在纷繁复杂的开源模型生态中，持续筛选出最优方案，确保底层技术的可靠性与前沿性。         在张策与珀西·梁的领衔下，他们与公司首席科学家特里·陶（Tri Dao），共同撑起了公司的底层竞争力，合力将Together AI的“算力成本”实现指数级下降，并从多个维度夯实Together AI的底层算法与系统能力，让公司在算法迭代、模型选型、系统架构、算力调度等方面持续领跑。

六 |          英伟达的“阳谋”：扶持生态卖更多GPU          再来看英伟达的出手，实际上，背后也暗含巨头布局AI算力生态的渴求。作为跟投方，英伟达不是第一次入局Together AI。         早在2023年11月，A轮融资时，英伟达就参与了其早期融资。彼时，英伟达也作为跟投方，投资该公司，而本轮融资中，领投方为Kleiner Perkins（凯鹏华盈）和沙特阿美旗下的 Prosperity7。         英伟达之所以持续布局这家公司，目的也很明确。截至目前，被英伟达看中的Together AI，已不是一家简单的AI云服务商了，而是英伟达庞大算力帝国中不可或缺的“生态节点”。简单来说，就是通过投资，英伟达直接锁定了巨额的硬件订单。         披露显示，Together AI的商业模式是建立在英伟达的H100、H200乃至最新的Blackwell GPU之上。

七 | 而Together AI年预订额突破11.5亿美元之后，势必也将对海量的GPU集群有更刚性的需求。英伟达的注资，刚好为自家芯片的持续消耗“铺路”。

八 |          除此之外，英伟达也正试图通过这类“新云”（Neocloud）服务商，打破传统云计算巨头的垄断，构建起更为广泛的算力分销网络。这对于英伟达来说，相当于通过扶持Together AI等服务商，将中长尾的AI开发者和企业需求打包，进而转化为更稳定、更可预期的采购曲线。         当然，更深层次的原因在于，英伟达正在从“卖硬件”的角色向“分收入”转型，做AI基础设施的合伙人。7月1日，英伟达已正式推出AI Compute Partnership（AI 算力合作计划），并明确其“产业链共建伙伴”的定位。

九 | 这意味着，英伟达将不止卖硬件，还将提供信用支持和收益共享机制，让Together AI等这类企业无需承担巨额的前期资本支出，就能快速扩张算力规模。

十 |          在这一模式和新定位下，英伟达已和Together AI之外的AI云服务商，如Sharon AI、Firmus等建立合作，他们将共建基于NVIDIA DSX AI Factory架构的“AI工厂”。         最新一季财报显示，英伟达交出了一份令人瞩目的成绩单：单季营收高达816.15亿美元，同比大增85%；归母净利润更是同比飙升211%，达到583.21亿美元。

十一 |          这一数据不仅刷新了英伟达季度营收的纪录，还助力英伟达站上5万亿美元市值，充分印证全球市场对AI算力基础设施的旺盛需求。

十二 | 而仔细翻阅这份绩单，不难发现，客户的多元化是此次英伟达业绩大增的一股力量之一。         财报显示，在数据中心业务中，英伟达的业绩增速需求，不仅来自超大规模云厂商（Hyperscaler），还广泛来源于AI云、工业、企业及主权AI客户，客户结构的分散化，降低了英伟达对单一大客户的依赖，让其构建起更具韧性的增长盘。同时，随着全球大模型迭代的加速，以及AI应用的快速落地，又进一步带动了全球AI工厂、数据中心的扩容，让下游云厂商、AI原生企业及政企的算力采购需求旺盛，增强了英伟达业绩向上的预期。         在英伟达的这场“阳谋”之下，Together AI也放话称，计划在未来五年内将公有云容量，也就是“算力”规模扩大50倍。这意味着，当算力成本不再是行业瓶颈时，属于开源大模型的时代或将真正到来，对于AI开发者而言，届时将是真正意义上的“Together AI”。

十三 |          而这家由华人技术力量主导的AI独角兽，也将乘着全球AI基础设施建设的东风，继续向前奔跑。

十四 |

Current article:http://www.zuanniangwandangliaolinechu.sbs/149xl/75brv.docx

Published on:01:35:25