新智元报道
大模型时代的下半场，全行业都在面对一道残酷的经济算术题。
随着 Agent技术的爆发，系统指令、历史对话、复杂工具调用被源源不断地塞入上下文。算力消耗呈指数级暴涨，但先进算力价格高昂且供给受限；而现阶段的普通算力，在显存带宽等物理指标上又面临着天然的瓶颈。
算力成本如果降不下来，AI 就永远无法成为赋能千行百业的「水电煤」。
破局点在哪里？
10月31日，在2026 GAI+C（全球人工智能与云计算）大会上，趋境科技携手生态共建伙伴智谱和摩尔线程，联合发布了「国产异构推理高品质 AI Token 工厂」解决方案。
趋境科技、智谱、摩尔线程共建人工智能国产生态
这是一次对产业痛点的精准狙击。
伴随方案浮出水面的，是资本市场极其敏锐的重注：近日，趋境科技正式宣布完成 B 轮融资，本轮募资超 13 亿元人民币。
至此，这家在 AI Infra 赛道深耕的独角兽，今年内已累计募资高达 24 亿元，高品质 AI Token 日均产量等核心经营指标在半年内实现了超十倍速的暴涨。
这套撬动数十亿资本的方案，究竟解决了什么产业难题？
它是如何在真实的生产线上，把性能特征各异的普通算力练成「Token 兵工厂」的？
要理解这套方案的产业价值，必须将其置于当下北京乃至全国人工智能产业基建的宏大语境中。
当前，北京正在全力打造全球 AI 创新策源地。海淀区作为核心基座，正在形成强大的产业生态。
北京期待在这个庞大的生态中，选择国内头部芯片公司以构筑「京芯」底层硬件，选择国内头部大模型以打造「京模」算法大脑；同时为弥合从「芯」到「模」之间横亘着的巨大的工程鸿沟，选择头部推理加速企业实现「京推」提效。
从「芯」到「模」之间，横亘着一条巨大的工程鸿沟：底层硬件的碎片化、普通算力显存带宽的物理限制，与上层大模型日益复杂的长文本推理需求之间，存在着严重的错位。如果只是浅做兼容和适配，根本无法满足真实的生产需求。
趋境科技勇闯难关，此次发布的方案，正是补齐了这块最难啃的关键拼图——「京推提效」。
通过全系统异构协同实现推理加速，不仅在宏观技术层面上取得了重要突破，更在北京海淀区打造AI创新策源地的区域背景下，精准填补了「京推提效」这一缺失的关键板块，实现了产业生态的闭环。
趋境科技创始人兼 CEO 艾智远主题演讲
在北京构建「京芯筑基、京模赋智、京推提效」的产业生态过程中，作为被身负厚望的头部推理企业，趋境科技从系统架构层面重构了普通算力的计算模式，使得产业闭环得以实现。
趋境科技在业界首创的全系统异构协同推理技术，不再只关注某一种 GPU，而是将集群中的 CPU、GPU、内存、网络全都纳入推理过程。
方案直指产业核心痛点，提出了全球首创三项创新技术：国产 Prefill-Decode（PD）异构协同、高性能异构 KV Cache 转换、异构算力计算融合。
这套系统级异构优化的核心产业价值，在于「Token 经济性」的全面跃升。
在同等资金投入下，该方案能够产出更多有效 Token，显著提升单位算力的生产效率；在同等产出下，又能大幅降低高品质AI Token 服务的整体交付成本。
换言之，该方案能够实现企业算力投入价值的数倍释放，让数百万级的硬件支出，承载起翻倍乃至更大规模的业务 Token 需求。
根据实测，这套国产异构方案的整体产出性价比，已经超越了采用清一色先进算力的基线方案。
它不仅让每一元算力投入转化为更多可交付的智能服务，更让国产旗舰模型能够在普通算力上，稳定输出复杂的长文理解与代码生成能力，真正服务全国产业，链接全球生态。
然而，从实验室走向真实的机房产线，中间是一片工程「死亡谷」。
我们与趋境科技首席架构师、开源项目 KTransformers 核心开发者谢威宇进行了深度对话，他对以下问题进行了详细解读。
趋境科技首席架构师、开源项目 KTransformers 核心开发者谢威宇主旨演讲
大模型推理底层包含两个阶段：Prefill（预填充）和 Decode（解码）。
Agent 时代，一个 100K 的长文本请求，光是中间状态（KV Cache）就能高达 6.54–8.87 GiB。
面对数倍的显存带宽物理差距，单一普通算力很难独立扛下顶尖模型的解码任务。趋境给出的答案极其直接：异构分离。 以某万亿级参数模型为例，趋境对计算强度进行了整体拆分。
让适合高计算量的普通算力负责 Prefill，让高带宽的先进算力（或未来的超级节点）负责 Decode，中间通过高速网络接力。这种方式让模型生成速度从原来的每秒 20-30 Token，直接翻倍跃升至每秒 50-60 Token。
但在真实的生产线上，把几十台不同架构的卡缝合在一起，无异于在悬崖边走钢丝。
「在实验室里完成一次跨芯片跑通，仅仅是工程的起点。」
趋境科技向我们展示了一份来自真实流量的震撼答卷：为了扛住平均 110K、最长 200K 上下文的真实请求，且满足首 Token 时延（TTFT）P50 小于 6 秒的极严苛标准，趋境采用了一种近乎「群狼战术」的配比。
对于多家国内算力，采用多台普通算力包围单台先进算力的阵型，使数据中心内绝大多数算力由普通算力承担；对于性能更强的的新一代国产硬件，以及在 AI Coding 等特定高并发场景下，则将配比进一步优化到极致。
不管底层用了几十台什么卡，最终交付给业务端的，必须是统一的高品质服务。
为了做到这一点，趋境必须做到以下三点。
第一层：必须「跑对」（精度无损）
两张不同架构的卡交接，权重差异、量化参数千差万别。稍有错位，模型就会悄悄「胡言乱语」。趋境从张量数值、短语义到长上下文逐层剥开验证，确保在 GPQA、AIME 等五项权威评测中，模型智商「一分没掉」。
第二层：必须「跑得动」（极致组网）
普通算力吃不下长文本，趋境通过软件把几十张小卡逻辑缝合成「一张大卡」。通过极其精密的手术刀级微批次编排，将跨卡组网的损耗死死压在 10% 以内。
配合 800Gbps 的标准 RoCE 网络，仅在两阶段交接的瞬间传输数据，彻底打通算力孤岛。
第三层：必须「跑得久」（容错自愈）
这是系统工程中最残酷的一环。
「拔掉电源。重启。一周，三四次。」
谢威宇坦言，在长期、满负荷的真实产线上，底层的 ECC 错误（显存物理读写错误）像幽灵一样防不胜防。面对这种硬件级的物理瘫痪，单纯靠算法无法修补。
「最难解决的，不是要求芯片厂立刻把硬件迭代好，而是我们外部这整套系统，怎么去『包容』这些随时可能出问题的卡。」
趋境构建了一套冷酷无情的容错防护网：建立轻量级的 Consensus机制。某台机器硬件报错瞬间瘫痪？系统会在毫秒级进行状态汇聚、异常隔离，流量瞬间平滑转移，故障节点原地重启。
底层的机房里可能正在经历硬件崩溃，但上层客户看到的，永远是一套行云流水、永不宕机的生产级服务。
工程上的极致压榨，最终转化为了商业数据上的狂飙。
面对市场对 Token 的极端品质分层，趋境放弃了兼容几百个模型，坚定选择「少模型、深优化、高品质」战略。针对 AI 场景需求做垂直深优化，因为企业级客户最终买单的，是高智能、低时延、强稳定的业务结果。
过去半年，趋境的整体业务已全面跨过盈亏成本线，实现了良性的自我造血。
现在，趋境正在参与构建多地、多源的算力供应链网络：立足北京的创新策源，公司以河南、杭州、海南为重点，结合区域资源与产业需求，加快高品质 AI Token 生产与服务网络建设，稳步推进联合运营合作。
目前，北京等地已建成多个工厂，其中一个数千卡规模的异构集群已上量，成功支撑了某大模型日均万亿级的高品质AI Token 供应。
不仅如此，趋境正在筹建两个万卡级异构集群，建成后将能支撑每日数十万亿的高品质AI Token 供应，满足国内每天几十到百万亿的巨大缺口。
那些沉睡在机房里的普通算力，正在被转化为源源不断的高品质 AI Token 供给。
随着芯片能力逐渐变强，PD 异构会不会只是一种过渡？
对此，谢威宇给出清醒的回答。
他将普通算力进入大模型生产比作「四级台阶」，而最后一级，就是 Prefill 和 Decode 全流程纯国产化。
今年下半年，随着新一代普通算力在显存带宽上的发力，趋境已经在实验室里看到了曙光。「预计到明年年中前后，我们就能看到大规模普通算力的全流程推理投产。」
届时，Decode 节点将仅仅依靠两种普通算力就能完成共同推理一个超大模型产出高品质AI Token 的异构时代。
更重要的是，异构协同的思想永远不会过时。无论是未来的多模态模型，还是硬件底层架构的演进，只要模型结构是异构的，这一最高效率法则就永远成立。
回头看那个一周三四次报错的故事，它带着几分狼狈。但这，恰恰是中国算力走进真实生产线必须交的学费。
「算力更高效，智能更普惠」，趋境通过持续的技术创新将这个愿景落实到了真实世界。
从北京海淀出发，中国的头部企业们正在构筑起「国模、国芯、国推」的中国人工智能基础设施的有效闭环，并将在不远的未来会把高品质的 AI 服务延伸至全球。
此时此刻，某台国产服务器的底层可能又报了一个错。但在数千公里外，客户的屏幕连一丝闪烁都没有，高品质的 AI Token 依然丝滑流出。
中国的人工智能基础设施，正在完成一场涅槃。
编辑：Aeneas 所罗门
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
In the second half of the LLM era, the whole industry faced a brutal economic arithmetical problem.
With the outbreak of Agent technology, systematic commands, historical dialogue, complex tool call have been constantly thrust into context. The increase in computing power consumption is exponential, but advanced computing is expensive and supply is limited; at this stage, ordinary computing is faced with natural bottlenecks in physical indicators such as VRAM bandwidth.
If the cost of calculating power does not fall, AI will never be able to become a “hydrocoal” capable of doing everything.
Where's the fucking point?
On October 31st, at the 2026 GAI+C (Global Artificial Intelligence and Cloud Calculus) Congress, Consistency Science and Technology, Together for Eco-Construction, Zhipu and Moor Lines, united publish to provide a solution to the "High-Quality AI Token Factory of National Isomer Logic".
Convergence Technology, Zhipu, Moor Lines for Artificial Smart Nations
It was an accurate sniper attack on the site.
Accompanying the programme is the extremely sensitive focus of capital markets: in recent days, convergence technology has officially announced the completion of the B round of financing, which is raising more than 1.3 billion yuan yuan.
So, this unicorn, who tilled deep on the AI Infra track, has raised a cumulative amount of $2.4 billion this year, and core performance indicators such as high-quality AI Token average daily production have increased by more than 10 times in six months.
What kind of industrial problem does this multibillion-dollar scheme solve?
How does it work on a real production line with ordinary calculations with different performance characteristics into "Token Armour Plant"?
In order to understand the industrial value of the package, it must be placed in an ambitious context in the current Beijing and national infrastructure for the AI industry.
Currently, Beijing is doing its best to create a global AI source of innovation. As a core base, the Seam Zone is creating a strong industrial ecology.
In this vast ecology, Beijing looks forward to choosing the domestic head chip company to build the base hardware of the "Kyochi" and the country's head LLM to build the "Kyomo" algorithm brain; at the same time, to bridge the huge engineering gap between the "core" and "model" by choosing the head reasoning to accelerate the firm's "Kyoyoku" effectiveness.
There is a huge engineering gap between "core" and "models": there is a serious mismatch between the fragmentation of bottom hardware, the physical limitations of the common calculus VRAM bandwidth and the increasingly complex long text reasoning needs of the upper LLM. If it is simple and compatible, real production needs cannot be met.
The publish project is the key piece of the puzzle, the "Kyoto Pro-Efficiencies", which is the most difficult to eat.
The acceleration of reasoning through system-wide isomer synergies has not only led to important breakthroughs at the macro-technical level, but also, in the regional context of the creation of the AI-based innovative resource base in the Beijing Haidian region, precisely filled the missing key block of "Kyoto-improvement " and closed the industrial ecology.
Founder and CEO of Treasure Technology.
During the construction of the industrial ecology in Beijing, "Kyochi's foundation, Kyo Mo's resourcefulness, and its effectiveness", Convergence Technology, as a head-debate firm with high expectations, has recast ordinary computing models at the structural level of the system, thus enabling industrial closure.
The system-wide synergetic reasoning techniques pioneered by convergence technology in industry no longer focus solely on one type of GPU, but incorporate all CPU, GPU, memory, network in clusters.
The programme refers directly to the industrial core and proposes the first three innovative technologies in the world: the domestic-generated Prefil-Decode (PD) isomer synergy, the high performance isomeric KV Cache conversion, and the integration of isomer algorithms.
The core industrial value of this system-level isomeric optimization lies in the overall rise in "token economics".
With equal funding, the programme is able to produce more effective Token, significantly improving unit-calculated production efficiency, and significantly reducing the overall cost of high-quality AI Token services.
In other words, the programme can release several times the value of the firm ' s calculus inputs, allowing millions of levels of hardware to be spent, to carry up to double or even larger business demands.
Based on actual estimates, the overall output value of the national isomeric programme has exceeded the baseline programme of using a clean and advanced calculus.
It not only transforms each meta-input into more deliverable smart services, but also enables the national flagship model to operate in ordinary computing power, stabilizes the ability to produce complex longform understanding and code generation, truly serves national industries and links global ecology.
However, from the laboratory to the real machine property line, there is a piece of engineering, Death Valley.
We engaged in an in-depth dialogue with the lead architect of Convergence Science and Technology and the core developer of the KTransformers project, Shewewu, who provided a detailed reading of the following issues.
Major Architect of Convergence Technology, KTransformers Core Developer of KTransformers
The bottom stage of LLM's reasoning consists of two stages: Prefil (prefill) and Decode.
In the Age of Agent, a 100K long text request can be as high as 6.54-8.87 GiB for KV Cache alone.
In the face of multiple VRAM bandwidth physical gaps, it is difficult for a single ordinary calculus to perform the decode task of the top model on its own. The answer given by convergence is extremely straightforward: isomer separation. For example, using a model of some trillion-degree parameters, the trend is to calculate the total strength of split.
Prefil, put high bandwidth, advanced calculator (or future super node) in charge of Decode, with a high-speed network relay. This way the model is generated at a direct doubling rate from 20-30 token per second to 50-60 token per second.
But on the real production line, the stitching of dozens of different structures together is tantamount to the taking of wires by the cliff.
"The completion of a cross-chip run in the laboratory is just the beginning of the project."
Convergence technology shows us a powerful response from real flows: In order to hold on to the true request of an average of 110K, up to 200K context, and to meet the very harsh criteria of the first Token latency (TTFT) P50 of less than 6 seconds, the trend has been to adopt a ratio that is close to the "wolf tactics".
For multiple domestic calculations, the use of multiple general-calculations arrays to encirclement single-stage advanced-calculations allows the vast majority of the computing power in the data centre to be absorbed by ordinary-calculations; for a more performing new generation of national-product hardware, and in a particular high-coding scenario such as AI Coding, the matching will be further optimized.
Regardless of what dozens of cards are used at the bottom, those ultimately delivered to the end of the business must be a unified high-quality service.
In order to do so, it is necessary to do the following three things.
First floor: "Run to Right" (no loss of precision)
The two different frames of cards are interfaced, with differences in weight and quantization parameters. A little bit wrong, and the model will whisper "buzz." The trend is to de-certify the scale, the phrase and the long context layer by layer, to ensure that the model's IQ is “mindless” in five authoritative assessments such as GPQA and AME.
Second floor: must run.
The average calculator can't eat long text, and there is a tendency to sew dozens of card logics into "a big card" through software. It's a very sophisticated surgical micro batch that puts the loss and death of cross-card network within 10%.
Complementing the 800 Gbps standard RoCE network, data are transmitted only at two-stage transversal intervals, with a complete calculation of the islands.
Level three: Must run long.
It's the cruel part of the system engineering.
Pull the power. Restart. Three or four times a week.
Shewyu says that on the long, full-load real line, the bottom ECC error is like a ghost. The physical paralysis of this hardware level cannot be repaired by algorithms alone.
"The hardest thing to solve is not to ask the chip factory to replace the hardware immediately, but how to get to the system outside of us, which could be problematic cards."
The trend has created a cold-blooded fault-resistant network: a lightweight Consensus mechanism. Some piece of machine and hardware just collapsed? The system aggregates at milliseconds, is abnormally isolated, the flow shifts smoothly and the failure nodes restart in situ.
The bottom room may be experiencing a hardware collapse, but the top customer sees a never-ending production-grade service.
The extreme pressure on the project eventually turned into a surge in commercial data.
Faced with the extreme quality layer of the market for Token, the trend has been to abandon hundreds of compatible models and to firmly choose a strategy of "less models, deep optimization, high quality". The vertically fine-tuning of the AI scenario, since the end of the payout for business-level clients is high intelligence, low latency, strong and stable business results.
Over the past six months, the corporate business of convergence has fully crossed the profit/loss line and achieved benign self-production.
The convergence is now engaged in building a multi-site, multi-source network of digital supply chains: a Beijing-based source of innovation, with companies focusing on Henan, Hangzhou and Hainan, accelerating the development of high-quality AI Token production and services networks, and steadily advancing joint business cooperation.
Currently, a number of factories have been built, including in Beijing, and one of the thousands of card-scale isomer clusters has been upscaled, successfully supporting a high-quality AI Token supply of a LLM average of trillions a day.
Moreover, the convergence is preparing two million-calories-class isomer clusters, which will support billions of dollars of high-quality AI Token supplies per day and meet the huge gap of billions to millions of dollars per day in the country.
Those ordinary calculations that sleep in the machine room are being converted into a constant supply of high-quality AI Token.
As the chip grows stronger, will the PD isomer be a transition?
We'll get a sober answer.
He compares normal algorithms to LLM production, and the last level, Prefil and Decode, is pure nationalization.
In the second half of this year, with the power of a new generation of ordinary algorithms on the VRAM bandwidth, convergence has already seen the dawn in the laboratory. "It is expected that by mid-year next year we will be able to see the full process reasoning of large-scale ordinary computing."
By that time, Decode node will be able to complete a joint inference of a high-quality AI Token era of LLM output with only two ordinary calculations.
More importantly, the idea of isomeral synergy will never be obsolete. Whether the future multi-modular model or the evolution of the hardware sub-structure, this maximum efficiency method is in place as long as the model structure is anisomer.
Turning back to the story of the wrong-doing three, four times a week, it's a bit of a mess. But it is precisely the tuition that China must pay to get into the real production line.
"Currently more efficient, intelligent and inclusive" is a vision that is being translated into the real world through continuous technological innovation.
From Beijing, China’s leading firms are building an effective shut-down of China’s artificial intelligence infrastructure – the model, the core, and the push – and will extend high-quality AI services globally in the near future.
At this point in time, the bottom of a national server may be reporting another error. But thousands of kilometres away, the client's screen is not even blinking, and the high-quality AI Token continues to slip out.
China ' s artificial intelligence infrastructure is completing a nitro.
Editor: Aeneas Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.