Chips

亚马逊Trainium芯片不止能运行、还能训练AI模型,10月29日财报前考验英伟达

SemiAnalysis实测显示,Trainium2与Trainium3如今已能在原生PyTorch下承担研究级训练;报告发布当天,安迪·贾西称推理成本仍贵了10倍,并把亚马逊自研芯片当作解法。这次松动的,正是那道把研究人员留在英伟达阵营的软件壁垒。亚马逊将于10月29日发布财报。

阅读原文

Richard TangRichard Tang · 4 min read
Share
2026年6月25日,印度新德里,亚马逊首席执行官安迪·贾西与印度总理纳伦德拉·莫迪握手。
1 / 5Slide 1 of 5
2026年6月,亚马逊首席执行官安迪·贾西与印度总理莫迪在新德里会面。贾西称,Trainium的性价比比同类处理器高出30%至40%。

亚马逊−0.72% — 亚马逊, down 0.72 percent today自研的AI芯片Trainium跨过了定义它的那条界线:现在,它不仅能运行模型,还能训练模型。独立芯片研究机构SemiAnalysis刚在Trainium3机架上实测数周,发布报告称,多数研究人员已经在用的PyTorch框架如今可以在Trainium上原生运行:代码能在动态图模式(eager mode)下逐行执行,正是研究人员所习惯的方式;提速工具torch.compile可以正常使用;几乎唯一要做的改动,是把`.to("cuda")`换成`.to("neuron")`1。

局限也说得直白:这套软件栈“与CUDA生态的成熟度相去甚远”——CUDA即英伟达−0.54% — 英伟达, down 0.54 percent today的软件层;测试中最大的模型为320亿参数,属于研究级规模,还算不上一次前沿模型训练1。

同一天,亚马逊首席执行官安迪·贾西把话题放在了需求端。他对风投机构a16z表示,推理——模型每次作答都要付费购买的那部分算力——成本仍然贵了一个数量级;他还称,Trainium的性价比比同类处理器高出30%至40%2。英伟达的训练定价,正是建立在SemiAnalysis刚刚量出的这层软件摩擦之上。而这层摩擦,正在消退。

切换成本曾是那道护城河

Anthropic$350B — Anthropic, private, latest valuation $350B是Claude背后的AI实验室,也是Trainium的头号客户,目前已在Trainium2和Trainium3上承担生产级推理和强化学习rollout1。迟迟未攻下的是研究训练:研究人员每天都在重塑模型架构,而旧的Trainium路径要经过XLA式图编译器,从CUDA迁移过来的研究者不得不重学训练循环1。

SemiAnalysis训练了开放权重的Qwen3-32B,修改了它的架构代码,还把gpt-oss训练配方——OpenAI$1.18T — OpenAI, private, latest valuation $1.18T开放权重的训练方案——直接开箱跑通1。结果:编译模式下达到每秒14,156个token、MFU为30.47%,较动态图模式下的每秒11,099个、MFU为23.89%提升1.28倍1。MFU衡量一次训练实际用上芯片峰值算力的比例。

在这背后,是一套规模最高达8万个物理核心的自动化质检体系:它对照英伟达基准,在研究人员共享代码的仓库Hugging Face上,对排名前10万的模型逐一核对数值1。AWS还聘请了曾主导谷歌在TPU芯片上原生运行PyTorch的工程师,在这里如法炮制1。每通过一项测试,研究人员就多一个无需改写代码即可迁移的模型。

利润率的证据先行一步

资金先于测试而动。2026年第二季度,即便资本开支持续攀升,AWS的营业利润率仍达到39.4%3。其签约在手订单达4960亿美元,2026年产能实际上已售罄,2027年产能也基本订满3。

Anthropic与OpenAI已披露的未来Trainium采购承诺合计超过2250亿美元2。本月早些时候,我们曾报道亚马逊出售80亿美元英伟达芯片的风险;那笔交易缺失的另一半,是Trainium能凭自家软件赢得业务的证明,而这一证明本周公之于众2。

亚马逊季度资本开支三年增长近四倍

$0B$20B$40B$60B2023Q42024Q22024Q42025Q22025Q42026Q22026年资本开支计划上调至2200亿美元
Data
资本开支
2023Q3$11.3B
2023Q4$13.35B
2024Q1$13.94B
2024Q2$16.39B
2024Q3$21.28B
2024Q4$26.05B
2025Q1$24.26B
2025Q2$31.37B
2025Q3$34.23B
2025Q4$38.47B
2026Q1$43.23B
2026Q2$53.08B
季度资本开支数据来自亚马逊向美国证券交易委员会(SEC)提交的文件;2026年全年约2200亿美元的计划为公司指引,并非图中绘制的数据。3,4

瓶颈转移至编译器

如今限制Trainium的,是软件打磨,而亚马逊正在花钱补齐这块短板。编译gpt-oss训练配方耗时约27分钟,对每小时都在迭代的研究人员而言纯属空耗;AWS表示,重构后的编译器目标是把这段等待最多压缩到原来的百分之一1。

亚马逊还接近以70亿美元收购Decart——这家以色列初创公司的优化工具栈,能把耗时数月的CUDA移植压缩到数周5。Anthropic也在对冲,自建了芯片团队,但量产芯片最早要到2028年才能问世6。对工程师来说,最值钱的技能正从编写CUDA,转向让模型在成本最低的芯片上高效运行。

10月29日:这笔交易将重新定价

确凿的事实,止于训练测试和AWS的利润率;再往后便是推测:更便宜的训练叠加更便宜的推理,将开辟贾西所说的“哑铃中段”——即他预计将成为AI最大需求来源的日常商业软件2。

亚马逊将于10月29日发布第三季度财报,预测市场Kalshi上的交易者认为,Trainium主导此次财报电话会议的概率为96%2。对亚马逊而言,问题是投资者是否将不再把自研芯片视为副业。对英伟达而言,问题则在于:训练是它最赚钱的业务,这块业务是否会开始对标一款芯片定价——而研究人员迁往这款芯片时,无需重写代码。

Get Chips by email, free

How AI demand is reshaping chipmakers, memory and the fabs behind them. In our daily email, every weekday.

Loading AMZN…

Reader comments

0 comments

    Chips

    Bank of America doubles Marvell's credit line to $3 billion as AI shorts count $6 billion of paper profits

    An 8-K filed October 23 shows a Bank of America-led syndicate committing five years of undrawn credit to Marvell at 0.875–1.5% over SOFR, double the old line and expandable to $5 billion, with a covenant drafted for a $500 million-plus deal. The AI financing-freeze short case now has to explain that page.

    Chips

    Murata turns away customers with record JPY 673.9 billion orders; October 30 capex decides

    Murata has record orders, a plant filling six years early, and customers it must decline. The capex number due with October 30 results decides whether that rationing lasts into 2029, and the ceramic powder, not the stock price, is the ceiling until 2028.

    Chips

    Fabrinet's CEO drew $13.6 million, 1,317 times its median worker, as the stock sits 57% off its high

    Fabrinet's new proxy shows a sixth straight record year, a $13.6 million CEO pay package at 1,317 times median employee pay, and just $5.2 million of buybacks. Shareholders reading it are holding a stock 57% below its high, ahead of November 2 earnings and a December 10 say-on-pay vote.

    Chips

    Why Wall Street Got the Ultra Clean Sell-Off Backward

    Ultra Clean fell 15% when Lam Research warned of supply bottlenecks, creating an inverted sell-off in the subsystem maker that builds the bottlenecked parts.

    Chips

    Nvidia backed its own AI customers with more than $170 billion; 81% run GPUs at half capacity

    The Information counts more than $170 billion of Nvidia commitments behind its own customers, Huang says only working capital is missing, and 81% of production AI operators run their GPUs at half capacity. The market spent Thursday pricing who pays.

    Chips

    United Microelectronics CFO's stake fell 28.4% in September; insiders sold $526 million and bought nothing

    The foundry's Form 6-K, filed October 23, 2026, shows chief financial officer Chitung Liu's stake fell 28.4% in September, inside a twelve-month stretch where insiders sold $526.4 million and bought nothing. On October 28 the market finds out whether the AI-power story can carry a multiple already 52% above its own five-year median.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.