首页 > 新闻 > 正文

小丈夫

WAIC2026：AI焦虑笼罩黄浦江_我的网站

新闻今日谈

一 | AUCKLAND, New Zealand -- Conservative former businessman Christopher Luxon will be New Zealand’s next prime minister after winning a decisive election victory Saturday.People voted for change after six years of a liberal government led for most of that time by Jacinda Ardern.The exact makeup of Luxon’s government is still to be determined as ballots continued to be counted.Luxon arrived to rapturous applause at an event in Auckland. He was joined on stage by his wife, Amanda, and their children, William and Olivia. He said he was humbled by the victory and couldn't wait to get stuck in to his new job. He thanked people from across the country.“You have reached for hope and you have voted for change,” he said.Supporters chanted his campaign slogan which promised to get the country “back on track.” Outgoing Prime Minister Chris Hipkins, who spent just nine months in the top job after taking over from Ardern in January, told supporters late Saturday he had called Luxon to concede.Hipkins said it wasn’t the result he wanted.“But I want you to be proud of what we achieved over the last six years,” he told supporters at an event in Wellington.Ardern unexpectedly stepped down as prime minister in January, saying she no longer had “enough in the tank” to do the job justice. She won the last election in a landslide, but her popularity waned as people got tired of COVID-19 restrictions and inflation threatened the economy.Her departure left Hipkins, 45, to take over as leader. He had previously served as education minister and led the response to the coronavirus pandemic. With most of the vote counted, Luxon’s National Party had about 40% of the vote. Under New Zealand’s proportional voting system, Luxon, 53, is expected to form an alliance with the libertarian ACT Party. Meanwhile, the Labour Party that Hipkins leads was getting only a little over 25% of the vote — about half the proportion it got in the last election under Ardern.And in a result that would be particularly stinging for Labour should it lose the seat, National was in a tight race for Ardern’s old electorate seat, Mount Albert. The seat has long been a Labour stronghold and was also held by another former Labour prime minister, Helen Clark. The National Party candidate for the seat, Melissa Lee, told The Associated Press she was feeling excited but also nervous about the final result in Mount Albert.“It’s been Labour since 1946. It has been the biggest, safest Labour seat forever,” she said. “It would be fantastic if we won it.” Lee said that when she was door-knocking, people had told her they were tired of the current government and were concerned with the state of the economy and the spiraling cost of living. David Farrar, a longtime conservative pollster, said there was still a good chance that Labour would end up holding the seat once all the votes were counted. However, he said, his initial impression of voting throughout the country was that it was turning out to be a “bloodbath” for the left. Luxon has promised tax cuts for middle-income earners and a crackdown on crime. Hipkins had promised free dental care for people younger than 30 and the removal of sales taxes on fruit and vegetables.Also at stake in the election is the government’s relationship with Indigenous Māori. Luxon has promised to axe the Māori Health Authority, which he says creates two separate health systems. Hipkins says he’s proud of such co-governance efforts and has accused Luxon of condoning racism.Within days of taking the reins in January, Hipkins found himself dealing with a crisis after deadly floods and then a cyclone hit New Zealand. He quickly jettisoned some of Ardern’s more contentious policies and promised a “back to basics” approach focused on tackling the spiraling cost of living. Warm spring weather in the largest city, Auckland, seemed to encourage voters, with queues forming outside some polling places. Early voting before Election Day was lower than in recent elections.During a six-week election campaign, both Hipkins and Luxon traveled the country and hammed it up for the cameras. Earlier in the week, Luxon, who served as chief executive of both Unilever Canada and Air New Zealand, told an energized crowd in Wellington that he would crack down on gangs.“I’ve gotta tell you, crime is out of control in this country,” Luxon said. “And we are going to restore law and order, and we are going to restore personal responsibility.”Luxon also got cheers when he promised to fix the capital’s gridlocked traffic with a new tunnel project.Luxon is relatively new to politics but held his own against the more experienced Hipkins during televised debates, according to political observers. But Luxon also made some gaffes, such as when he was asked in a 1News debate how much he spent each week on food.His answer of “about sixty bucks” (U.S. $36) was ridiculed on social media as showing he was out of touch with the cost of living. ___Follow AP’s Asia-Pacific coverage at https://apnews.com/hub/asia-pacific。 2026 WAIC世博中心南方周末研究员/摄原价168元的单日票，黄牛和二手平台售价1000-1500元不等，三日通票卖到3000元。展位租赁费2880元/㎡，36㎡起订，加上设计、物料、装修等费项，展台支出高达百万元…… 2026年7月17-20日，WAIC 2026盛况空前，巨额费用没有吓退观众和参展企业。在连日37度高温的上海，AI比天气更滚烫，刷新了WAIC往届全部参展和观展纪录。

二 | 根据主办方的数据，在4天展会里，超过40万人次涌入展馆，全球线上流量累计超过30亿；展览面积首次突破10万平方米，1117家企业带来了4486项展品，其中有351款是全球首发新品，连伊利这类看似与AI关联不大的公司都要展示一番“AI牛奶”。不止于此。

三 | 以东浩兰生展览馆为圆心，向外延展到徐汇、张江、前滩……超百场各种主题的Afterparty占领了周边的报告厅、咖啡厅、餐吧、酒馆、酒店、游艇，甚至是洗浴中心。

四 | 白天的展会结束后，技术交流、投资洽谈、人才招聘等活动换个场地，再续杯。为了抓住资金与聚光灯，多数公司提前数日做预热与新品发布，把WAIC的战线拉长到十余天。黄浦江畔织起AI赛博梦，也弥漫着技术焦虑、投资焦虑、人才焦虑和商业化焦虑。

五 | 具身智能、大模型、AI基建，展馆每个角落都在宣告“未来已来”，但真正的问题藏在热闹背后——这些机器人、模型和算力，距离场景应用和落地还有多远？具身赛道拥挤 2026年年初，一家机器人初创公司的创始人对南方周末表达了对“大厂下场碾压”的担忧。

六 | 现在，这个“噩梦”正在逼近。具身智能赛道越来越拥挤，继BAT（字节、阿里、腾讯）三家大厂之后，一些中厂和AI模型厂也先后入局。例如，蚂蚁灵波卡在半个月前，密集“日更”了一系列VA、VLA、世界模型和世界动作模型；WAIC开幕前三天，小米先是宣布了机器人进厂的进展，随后又甩出一款生成模型和一款具身基座模型；而京东直接将参展主题定为“AI进入物理世界”，其间还开源了行业最大的人类第一视角数据集EgoLive。在计算机视觉时代的“AI四小龙”中，商汤孵化的大晓机器人在展会期间发布了Kairos 3.1世界模型，并达成了一波具身零售落地合作；旷视团队再创业的原力灵机，现场展示了“机器人拼装长城模型”。在大语言模型时代的“AI六小龙”中，智谱和MiniMax此前均成立了与机器人相关的子公司。此外，端侧模型独角兽面壁智能也在WAIC期间正式宣布迈向具身智能领域，发布了系列新模型。 X-Era Lab机器人穿鞋带南方周末研究员/摄在切入点上，这些模型厂商往往基于自身视频生成模型的能力，顺势做世界模型；大厂则有场景优势，在数据集、全栈方面领先。没落的“AI四小龙”需要复苏，竞争胶着的“AI六小龙”也亟需寻找AI Coding之后的新的估值支撑，机器人或许会是一个不错的锚点。跨界的大中小厂，叠加上百家新成立的具身公司，尽管只有H3一个展馆，具身智能实际已成为WAIC期间最为拥挤的一个赛道。竞争对手虎视眈眈，机器人公司还要一边搞研究，一边做落地。这一方面是技术焦虑，展会期间及前后，新发布了十余个相关模型，新模型愈发强调对物理规律的理解，以及训练数据、架构等方面的“具身原生”性。创新大多在细枝末节上，比如数采方式、数据配比、VLA与世界模型的融合方式。各家都宣称自己在五花八门的评测榜单上击败了硅谷头部机器人公司Physical Intelligence（PI）的上一代模型，但对于刷榜“应试”的质疑也一直在，来自模型架构层面的、更触及本质的创新仍未被看到。

七 | 对模型突破的渴求，又催生了人才焦虑。

八 | 7月19日傍晚，蚂蚁灵波包场了展馆对面的一间咖啡厅，借着技术交流的名义，将上海各高校的青年学生们聚在了一起，多名HR直接在现场收简历。整个WAIC期间，这种以人才招聘为主要意图的开发者party不胜枚举。

九 | 另一方面，投资焦虑又催生了商业化焦虑。中国的创投环境仍相对谨慎，有投资人向南方周末调侃“我们吃不下太大的饼”，不做全栈的项目会被认为“想象力不足”，没有商业订单的项目更加无法容忍。 PI公司的研究科学家任至意在WAIC上说，PI更多是一个以研究为主的公司，寄望于先把AI基础设施做好、让模型能力涌现出来，短期内不会去考虑落地。

十 | 相比之下，这样的公司在中国很少见。大家都忙着在一个还没有打好的地基上考虑“卖楼”。

十一 | 在展览现场，几乎所有公司都展示了落地Demo，多样抓取、包裹分拣、物料搬运、商品打包是最常见的几个场景，机器人推理和干活的速度只有人类的0.3倍速甚至更低。少数动作流畅利落的机器人，那也并不必然与先进划等号——大概率是预设好的程序。

十二 | 这衍生出一种类似L2自动驾驶的商业化模式，即先把机器人部署进场景里，由人类遥操员远程兜底，担任与智驾安全员相似的角色，收集 Corner Case和Failure Case反哺模型训练。

十三 | 总之，技术焦虑导致在模型上刷榜，商业焦虑导致在落地上雕琢demo，这些又或多或少折射在融资与估值表现上。在这场长线竞争中，每家公司的身位都很模糊，投资人多是靠着直觉押注、跟投头部，或者广撒网、多敛鱼。机器人确实在变聪明，参展商展示的技能越来越长程和泛化，柔性和精细操作变多；为了证明模型的鲁棒性，现场刻意施加了更多干扰。只不过，这种进化的速度仍比较缓慢。 “具身智能可能不会出现像大语言模型‘ChatGPT时刻’那样清晰的拐点，而是会以一种润物无声的方式逐渐成长。

十四 | ”星海图合伙人、前高瓴资本投资人罗天奇认为。算力底座多元作为支撑AI场景落地和技术演进的关键底座，算力基础设施早已走向台前。H2展厅内，观众驻足在那些外观并不起眼、如同黑箱一般的机柜面前，举着手机排队拍照，好奇地询问“这个大家伙到底有什么用？” 华为昇腾950超节点南方周末研究员/摄无论机器人实时处理信息，还是智能终端调用模型，背后都依赖算力。在智能体时代，算力的衡量标准不再是堆卡，而是能否稳定、高效地产出Token。从单颗芯片到大规模集群，从服务器到高速互联和存储方案，算力正走向多元架构与系统能力的综合较量。华为的热度极高，首次线下展示昇腾950超节点（Atlas 950 SuperPoD），本质是将1024张加速卡通过高速互联协议整合至一台超级计算机，追求单系统的极致通信密度和内存一致性。多个超节点进一步横向互联，便构成了更大规模的智算集群。中科曙光带来了中国首个全国产十万卡AI超集群——曙光8000（登峰）真机，也是本届WAIC 的“镇馆之宝”。曙光8000现场只开放了部分区域，其余机柜以黑色挡板遮蔽。

十五 | 2026年7月10日这一AI超集群落成时，已同步接入国家超算互联网，进而接入全国一体化算力网。

十六 | 至此，中国的AI基础设施建设，从千卡、万卡级迈向十万卡级部署阶段。此外，摩尔线程、沐曦股份、壁仞科技、天数智芯、燧原科技等初创GPU公司，也都集中展示了超节点和算力集群。

十七 | 在从单卡走向集群的过程中，连接方式也悄然重构。例如，摩尔线程MTT C256超节点首创一层Scale-up网络，通信无需逐层转发，减少了带宽损耗和等待时间；燧原科技OEX超节点则采用正交无背板设计，将机柜内的线缆基本去除，减少互联成本和迭代周期。

十八 | 集群之外的另一大特点是GPU“第三条技术路线”。东方算芯展示了软件定义近存计算3D AI芯片DF1000，直接打出“14nm≈4nm”的口号。

十九 | DF1000通过底层计算架构创新，通过软件定义芯片技术实现软硬件解耦和动态重构，以空间并行的时分复用显著提升硬件资源的利用率；通过3D混合键合技术将计算层与存储层垂直堆叠，压缩互连间距至亚微米级，缓解“内存墙”“带宽墙”和“功耗墙”的瓶颈。有别于GPGPU和ASIC，东方算芯DF1000的技术源头来自2006年清华大学实验室，最初是面向超算场景设计，在这轮AI浪潮兴起时转向智算领域。

| 与之类似，清微智能主力芯片TX81采用的是可重构技术，兼具GPU的灵活性与ASIC的高效能，并基于此打造了REX81 Supernode超节点系统。 “发动机很重要，但用户买的是整台汽车。”清微智能副总裁黄文彬对南方周末表示。在他看来，芯片、系统、集群的协同创新，才能让AI真正跑起来。 REX81 Supernode超节点南方周末研究员/摄这种多元探索是产业走向成熟的必经阵痛，也是制程受限之下焦虑的投射。中国工程院院士、清华大学计算机系教授郑纬民坦言，国产卡的显存内存、容量、带宽、延迟与海外仍有差距，建立算力不等于高质量的Token产能。 “高质量Token意味着延迟很短，很值钱。

| 如果问两分钟才开始说话，每秒钟才为一个人服务，就不值钱。”郑纬民表示。如何将算力基建转化为高价值的服务？有效算力和算力性价比成为下一代算力集群的关键指标。一批Token工厂也顺势涌现，例如硅基流动、无问芯穹、九章云极等。如果说算力集群是“发电厂”，Token工厂就是“电力公司”，以智算集群为基础，经过调度、适配和优化，将算力转化为可计价、可调用的Token服务。这一模式是对当前算力错配的回应。

| 一方面，AI应用对算力的需求异常强烈，例如Kimi K3发布不到三天，用户的请求量便逼近其承载极限，导致公司暂停C端的新用户订阅；另一方面，大量算力集群在实际运转中却存在资源闲置，有些园区的算力利用率不足五成。九章云极的做法是用DCU“统一度量衡”。它发布了一套完整的AI规模化生产体系，将不同厂商芯片的算力统一折算，让客户能像购买电力一样按需购买算力。也就是说，把过去“整机租赁”的算力供给变成“按需取用、随用随付”的弹性服务。用系统创新弥补单卡短板，用工程能力兑现算力价值，其间技术追赶的紧迫、商业闭环的未知，都催促着产业不断试错。AI落地的叙事，终究要回到朴素的商业逻辑——谁能用更低的成本、更稳定的系统，产出更便宜的Token，谁就站住了身位。