文｜硅碳变量，作者｜胡一刀
在AI PC赛道热闹了两年之后，Windows阵营终于拿出了真正对标Mac生态的旗舰筹码。
微软联合英伟达推出搭载RTX Spark超级芯片的Surface Laptop Ultra，这不再是微软单独下场做硬件的一次常规迭代，而是Windows生态首次拿到英伟达这款一体化顶级SoC，依靠端侧大算力补齐长期存在的短板。
微软Windows与设备执行副总裁帕万·达武卢里周三在旧金山举行的一场公司活动上表示，Surface Laptop Ultra现已开放预购，并将于10月16日开始发货。该电脑配有磁吸USB-C端口，类似苹果的MagSafe充电接口。
Counterpoint Research此前预计人形机器人存储需求5年增加20倍，马斯克直接驳斥称其“荒谬”，并表示机器人所需的内存容量将会超过地球上其他所有设备的总和。
因为要让机器人依托物理基底，才能完成原子级塑形。超级智能的算力需求更是会远远超出这个量级，但几乎全部的超级智能算力都将部署在深空计算节点。
但二者的分歧从来不是关于DRAM需求量的数字博弈：一边是成熟调研体系，沿着已商业化终端的增长路径做线性外推；另一边则将具身智能的终点定义为物质重构，再从终点倒推基础设施的需求上限。
可见，两套体系讨论的本就不是同一时间尺度和同一类任务形态。
两套估算框架，错位的前提假设
分析师的测算框架，扎根于半导体行业沿用多年且广泛通行的一套估算范式。
它的逻辑链条并不复杂：以单机存储规格、终端出货量为核心，叠加渗透率假设，估算赛道整体需求。
自动驾驶，是这套框架最有参考价值的样本。量产辅助驾驶的任务范围相对可控，典型场景结构化程度高，算力、存储的搭配方案沉淀了充足的落地数据，可供横向参照。
这也使得过去一段时间，不少机构在测算人形机器人硬件增量时，习惯于参照车载算力的落地经验进行外推。
但这套推演建立在若干隐性前提之上，这些内容很少会在调研报告中写明。
其一，是任务范围的约束。短期商业化人形机器人落地场景集中在搬运、巡检、简易装配、人机协同这类标准化作业，配套的世界模型偏向轻量化和局部化，按需触发更新，无需持续维护一套全局和高精度的物理状态表征。
其二，是算力调度的分层设计：感知与基础规划在端侧完成，复杂模型推理和长周期任务调度回传云端。存储负载主要承载图像、点云、短时状态缓存以及策略权重，负载构成和车载系统较为接近。
其三，增长路径为渐进式释放：需求随良率优化、成本下行、场景渗透稳步抬升，它是半导体大盘之上的新增量，而非颠覆现有产业框架的范式变革。
这套推演并非空中楼阁，它和当前Optimus以及绝大多数人形机器人厂商的量产思路高度契合。想要控制BOM成本、推进商业化落地，工程师的选择往往是压缩板载内存和优先保障带宽，借助模型蒸馏降低硬件算力压力。
这里也暴露出这场讨论中最普遍的混淆：不少观点会把面向量产的短期硬件方案，与马斯克描绘的远期终局放在同一维度对比。
量产工程要回答的是如何稳定且低成本地交付产品，而那句“荒谬”的评价指向一套时间尺度和前提假设全然不同的智能愿景。
说到底，Counterpoint给出的是一份面向中期商业化周期的增量测算。
它预设人形机器人作为一类新型终端嵌入现有生产体系，任务复杂度可控且成本红线刚性，迭代节奏参照消费电子和汽车电子已验证的迭代路径。
报告并不预判硅基智能能否改造物质世界，只回答一个问题：如果人形机器人沿着当前路线落地，存储市场将迎来多大增量。
马斯克的质疑，则直接跳出了这套预设。他不再以现有终端形态为标尺，而是重新定义具身智能的终极目标：原子级塑形。
沿着这个设想，智能体需要持续维护高保真和强耦合的物理世界模型，追踪形变、力学反馈、摩擦、相变等物理过程，每一次抓取、接触、对物质的干预都会带动模型状态同步更新。
在这个前提下，决定存储需求体量的不再是大模型参数量，而是这套持续演化和彼此耦合的物理状态空间。
构想背后，难以绕开的工程硬约束
原子级塑形，并不是放大版的3D打印。
它指向一套闭环且可持续的物质操控能力：智能体需要维护高保真、强耦合的物理世界模型，实时追踪固体形变、接触力学、摩擦效应、热传导、材料相变等动态过程。
每一次抓取、组装、改写物质结构，都要同步更新规模庞大的状态空间。
传统大模型处理文本、图像这类符号化信息，数据支持分片、采样与压缩，一定程度的失真尚可接受；而面向原子尺度操控的具身智能，各个状态之间存在强因果关联，局部微小扰动会沿着物理链路持续传导扩散。
那句“内存容量超过地球上其他所有设备总和”，也经常被简化误读为单台机器人需要搭载海量存储。
更贴合原始表述逻辑的解读，指向集群协同场景：大量机器人不再独立执行零散指令，而是协同完成物质重塑，共享、同步一套全局物理世界模型，任意一次物理交互都会驱动全局状态同步刷新。
当然，支撑这套猜想成立的前置条件极为苛刻：全局状态空间是否支持分片、分层缓存？如果把高频局部状态留在端侧，低频全局推演交给远端算力，能否有效分摊存储压力？围绕这些问题，行业还没有成熟的工程落地路径，目前仍然只是一个有待验证的假说。
深空计算节点的构想，也不宜简单贴上科幻标签。
它本质上是一套试图突破地面算力天花板的突围思路，地面超算集群长期受限于多重刚性约束：散热极限、能源密度、重力扰动以及土地与电力成本。
一旦具身智能的目标从信息生成转向大规模物质改造，算力需求或将膨胀到地表基建难以承接的程度，向外寻找新的算力载体在逻辑上是一个值得探讨的方向。
算力选址的迁移脉络有迹可循：从分散机房走向大型数据中心再到海底、极地算力集群，背后共同的逻辑是寻找温度、能源、电磁干扰条件更优的部署环境。
深空计算节点，可视作这条演进线索被推至极限的一种猜想。但光速时延是这套构想很难绕开的约束，也是最核心的工程矛盾。
地表人形机器人的力控、触觉交互与精密装配，高度依赖毫秒级闭环反馈。地表与深空节点之间存在固定光程延迟，远端算力很难承接这类对实时性要求严苛的物理操控任务。
沿着这项约束推演，一套折中分工框架或许会慢慢成型：端侧机器人保留局部物理模型、实时感知与闭环控制，承接低时延刚需；深空超级节点负责长周期任务规划、全局物质系统优化、超大尺度世界模型推演以及远期路径搜索。
这套分工模式有望降低深空算力落地的门槛，但也可能同步弱化它的不可替代性。
如果绝大多数高价值的实时闭环控制都留在地表端侧，深空节点更接近于远期补充算力池，而非这套具身智能体系里不可缺少的核心。
除此之外，在轨芯片还要面对辐射加固、高昂发射成本、在轨维护困难、星间激光通信稳定性、存储介质在空间环境下的寿命衰减等一系列高门槛挑战。
它算不上一条随时可以启用的备选路线，只有当地面方案真正撞上实质性瓶颈之后，才有可能进入严肃的路线评估环节。
将这两层约束合并审视不难看到：马斯克描绘的图景是一套逻辑自洽的思想实验，但它的成立有赖一连串苛刻假设同时兑现。
只要任意一环工程瓶颈无法打通，这套关于存储与算力需求的推演逻辑就可能大幅收敛。
它可以成为长期技术探索的方向，却很难用来修正未来五年的产业需求曲线。
两套并行叙事，一场筹码分配游戏
存储产业链当下，并存着两套截然不同的叙事剧本，厂商与资本也在两种预期之间分配筹码。
第一套是温和放量的剧本，与Counterpoint的测算逻辑保持一致。
LPDRAM、高可靠存储与高带宽接口，将伴随人形机器人商业化渗透带来稳步增长的需求。这条路线的增量可量化，产能便于有序规划，二级市场也可以沿用成熟估值框架定价。
它的风险集中在商业化落地不及预期、场景拓展缓慢、终端需求迟迟无法起量，这也是当前产业资源投入的主线：存储厂商的扩产计划、存储接口标准迭代、配套测试方案开发，大多以这套叙事作为基准。
这条路线足够保守，风险边界清晰，更适合头部大厂与公开市场资金押注。
第二套是范式跃迁的剧本，对应马斯克提出的远期假设。
需求不再沿着现有曲线线性放大，而是伴随智能任务底层逻辑的切换，或将倒逼整套存储架构重构。近存计算、高带宽堆叠、持久内存、分布式共享状态池有望迎来全新落地场景，这条路线兑现概率偏低，但一旦跑通将重塑存储赛道内部的价值分配格局。
目前参与其中的更多是前沿实验室、小型研发团队以及一级市场风险资本，虽然尚无法主导当下的产能规划，但在潜移默化中抬升了部分前沿技术路线的研发优先级。
产业资本的布局策略也随之分化：把大部分筹码押注在确定性更强的中期增量上，同时拿出小部分资源提前预研适配超大物理状态空间的新型存储架构。
这套双线布局，本身并不矛盾。
Optimus团队持续压缩单机内存配置和优先扫清量产障碍，其长期研究小组也在推演面向物理世界模型的下一代硬件形态。短期工程落地优化与远期技术猜想，完全可以在同一主体内部共存。
分析师的预测本质上是对已有成熟范式做线性外推，这套方法擅长估算手机、服务器、车载电子带来的硬件增量：样本充足、变量收敛，历史数据的相关性能够为预测提供支撑。
可一旦技术目标发生底层切换，AI不再局限于感知和解读信息世界，转而走向对物质结构的主动改造，依托历史样本画出的增长曲线很可能失去参照基础。
马斯克的判断，并不是一份标准意义上的产业需求预测。它更像一份目标陈述：先锚定具身智能想要抵达的终局，再反向推导底层基础设施需要具备怎样的能力上限。
一套叙事向后看，依托历史规律推演增量；另一套向前锚定终点，倒推硬件体系的极限边界。
在未来很长一段时间里，存储行业都会悬浮在两套叙事中间，没有人能够笃定哪一条路径必然成型。产业只能踩着看得见的台阶稳步推进商业化，同时为那些遥远甚至最终无法落地的设想悄悄埋下技术伏笔。
分歧的核心从来不是5年20倍或是更高的需求量级，而是我们愿意为硅基智能划定怎样的能力边界。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The vinyl silicon carbon variable, by the author.
After two years of intense activity on the AI PC track, the Windows camp finally came up with a real flagship chip on Mac's ecology.
The introduction by Microsoft of Surface Laptop Ultra, which carries the RTX Spark superchips, is no longer a routine integrator of Microsoft's next hardware, but is the first time that the Windows ecology has been given an integrated top-level SoC in Britain, with long-standing shortboards that rely on side-calculation.
Microsoft Windows and Pavan Davululy, Executive Vice-President of Equipment, stated at a corporate event in San Francisco on Wednesday that Surface Laptop Ultra was now open for pre-purchase and would commence delivery on 16 October. The computer is equipped with a United StatesB-C port of magnetic ingestion, a MagSafe charging interface similar to an apple.
The Counterpoint Research previously predicted a 20-fold increase in human-shaped robot storage requirements over five years, which Mask directly refuted as “preposterous” and stated that robots would require more memory capacity than the sum of all other equipment on Earth.
Because it takes the robot to get to the physical base to complete the atomic form. Super-intellectual computational needs go far beyond this scale, but almost all super-intellectual computations will be deployed at deep-space nodes.
But the difference between the two is never a digital game about DRAM demand: on the one hand, there is a mature research system that pushes linearly along the growth path of a commercialized terminal; on the other hand, the smart endpoint is defined as a material re-constructing and then pushes the demand cap on infrastructure from the endpoint.
Thus, the two sets of systems were not discussed at the same time scale or in the same type of task.
Two sets of estimation frameworks, false assumptions
The framework for the measurement of analysts is rooted in a set of estimation models that have been used in the semiconductor industry for many years and are widely used.
Its logical chain is not complex: the total demand for the track is estimated using a single-machine storage specification, terminal output, superseding permeability assumptions.
Automatic driving is the most valuable reference sample in this framework. The scope of the task of mass-assisted driving is relatively manageable, the typical scene is highly structured, and the calculator-stored blending programme deposits sufficient landing data to be cross-referenced.
This has also led many institutions over the past time to be used to extrapolate in the light of the landing experience of vehicle-mounted computing when measuring the increase in human robotic hardware.
However, this exercise is based on a number of hidden premises, which are rarely documented in research reports.
First, there are constraints to the scope of the mandate. The short-term commercialized human robotic landing scene is concentrated in such standardized operations as handling, inspection, simple assembly, human synergy, with the associated world model being light-looking quantization and localized, triggering as-needed updating, without the need to maintain an ongoing set of global and high-precision physical state signs.
The second is the stratification of the algorithm: perception and base planning is done at the end, complex model reasoning and long-cycle tasks are sent back to the cloud. Stores the main loads of images, clouds, short-term status cache and strategic weights, which are made up of loads closer to the vehicle load system.
Third, the growth path is a gradual release: demand is rising steadily with good rates, down costs, and scene penetration, which is an addition to the semiconductor disk, rather than a paradigm change that undermines the existing industrial framework.
This is not an aeroplane, it is highly compatible with the current quantitative thinking of Optimus and the vast majority of man-made robots. In order to control BOM costs and drive commercial landings, engineers often choose to compress plate memory and preferentially secure bandwidth to reduce the pressure on hardware by model distillation.
Here, too, the most common confusion in this discussion is that a number of views will compare short-term, production-oriented hardware options with the long-term end of the Musk depiction.
The quantitative engineering answer is how to deliver the product in a stable and low-cost manner, and the “puzzling” evaluation points to a set of timescales and assumptions with completely different intelligent visions.
Ultimately, Counterpoint gave an incremental measure for the medium-term commercialization cycle.
It foresees human robotics as a new type of terminal embedded in the existing production system, with a manageable and cost-intensive red line of task, and an iterative rhythm based on an iterative path that is certified by consumer and automobile electronics.
The report does not prejudge whether silicon-based intelligence can transform the world of matter, but answers only one question: If human robots land along current routes, the storage market will receive much more.
Mask's question went straight out of the preset. Instead of using the current terminal form as a yardstick, he redefined the ultimate goal of intelligence: atomic-grade plastic.
Along this line, AI agent is required to continuously maintain high-level physical-world models of security and alignment, to track physical processes such as shape-formation, mechanical feedback, friction and transformation, with each capture, exposure and intervention in matter leading to a synchronized update of the state of the model.
In this context, it is no longer LLMparameter count that determines the volume of the storage demand, but rather the physical state space that continues to evolve and converge.
It's hard to get behind it.
Atomic-grade plastics, not 3D printing in magnified versions.
It points to a set of closed and sustainable material manipulations: AI agent requires the maintenance of high-level, well-coordinated physical world models to track, in real time, the dynamics of solid-forming, exposure mechanics, friction effects, heat transfer, material transformation, etc.
The large-scale state space is updated simultaneously every time a material is captured, assembled and rewritten.
Traditional LLM processes text, images, such symbolic information, data support fractions, sampling and compression, and some degree of deceptiveness are acceptable; however, there is a strong causal link between the various states, with local microturbations spreading continuously along the physical link.
The phrase “the memory capacity exceeds the sum of all the other equipment on Earth” is also often interpreted as being simplified to mean that a single robot needs to carry a mass storage.
A more accompanied interpretation of the original narrative logic points to a cluster synergetic scenario: a large number of robots no longer independently implement fragmented commands, but work together to remodel the material, share and synchronize a global physical world model, and any physical interaction will drive the global state to synchronize.
Of course, the pre-conditions that support this assumption are extremely harsh: does the global space support the fraction, the layer cache? If the local state of the HF is left on the end side, can the low-frequency global rollover be handed over to the remote calculus, and can the storage pressure be effectively shared? Around these issues, the industry has not yet matured the path to land and is still a hypothesis to be tested.
The concept of deep space nodes should not simply be labelled as sci-fi.
It is, in essence, a scrambling line of thought that seeks to break through the ground-based computing ceiling, where over-calculated clusters are subject to long-term constraints of multiple rigidities: heat dispersion limits, energy intensity, gravity disturbances and land and electricity costs.
Once the smart target has shifted from information generation to large-scale material modification, calculating power needs or expanding to the extent that surface infrastructure is difficult to sustain, it is logically a direction to look for new computing vehicles.
The migration routing of computational sites is consistent: moving from a decentralized room to a large data centre to a sub-seabed, polar-calculations cluster, with the common logic of looking for a better deployment environment for temperature, energy and electromagnetic interference.
Deep space calculation nodes can be considered as a conjecture that this evolutionary clue has been pushed to the limit. However, light speed latency is a constraint that is difficult to circumvent and is at the core of the project.
The surface man-shaped robot is highly dependent on msCyclical feedback for power control, touch interaction and precision assembly. There is a fixed radial delay between the surface and the deep space nodes, and it is difficult for far-end computing power to take on such physical manipulation tasks that are demanding in real time.
Following this constraint, a compromise division of labour framework may slowly take shape: end-end robots retain local physical models, real-time sensory and closed-ring controls, taking on low latency requirements; deep-space supernodes are responsible for long-cycle mission planning, global material system optimization, hyperscale world modelling and long-range routing.
This division of labour model is expected to lower the threshold of deep-space computing, but it may also be synchronized to weaken its non-alternativeity.
If the vast majority of high-value real-time closed-ring controls remain on the surface side, deep space nodes are closer to the long-term supplementary calculus than the indispensable core of the AI agent system.
In addition to this, the on-orbit chips face a series of high-threshold challenges, such as radiation reinforcement, high launch costs, difficulties in-orbit maintenance, stability of inter-planet laser communications, and the decay of the life of storage media in the space environment.
It is not an alternative route that can be activated at any time, and it will be possible to enter a serious route assessment only if ground programmes actually hit substantive bottlenecks.
It is not difficult to see these two layers combined: Mask painted a picture of a logical and self-conforming experiment of thought, but it was built on a series of harsh assumptions that went hand in hand.
As long as any one of the engineering bottlenecks cannot be broken, this evolutionary logic of storage and computing needs can be significantly reduced.
It could be the direction of long-term technological exploration, but it would be difficult to modify the industry demand curve for the next five years.
Two sets of parallel narratives, a chips game.
The chain of storage industries, with two distinct sets of narrative scripts, also distributes chips between the two expectations.
The first set is a mildly distributed script, consistent with the measurement logic of Counterpoint.
LPDRAM, high-reliability storage and high-bandwidth interfaces will be associated with the commercial penetration of human-shaped robots, which will generate steadily growing demand. The increase in this route can be quantization, which facilitates orderly planning of capacity, and the secondary market can follow the mature valuation framework.
Its risk is concentrated on commercialization that falls short of expectations, the slow expansion of scenes and delays in the start of terminal demand, which is the main line of current industrial resource input: storage plant expansion plans, storage interface standard iteratives, and supporting testing programme development, most of which are based on this narrative.
This route is sufficiently conservative, with clear risk boundaries, and better suited to bet on capital and open market funds.
The second set is a model leap script, which corresponds to the forward assumptions put forward by Mask.
Demand is no longer zooming in along the existing curve linearly, but is accompanied by a transition in the underlying logic of an intelligent mission, or will be re-engineered as a whole storage structure. Close counting, high-bandwidth stacking, long-lasting memory, distributed shared-state ponds are expected to be a new landing scene, a route that is less likely to materialize, but which, once running, will recreate the value distribution pattern within the course.
Currently, more people are involved in front-line laboratories, small R & D teams and first-level market venture capital, which, while not yet in a position to dominate current capacity planning, has raised some of the forward-technology development priorities in submersion.
Industrial capital deployment strategies have also been fragmented: much of the chips have been betped on more defined medium-term increments, while a small portion of resources have been used to pre-empt new storage structures for super-heavy physical space.
This double-wire layout is not in itself contradictory.
The Optimus team, which continuously compresses the single memory configuration and prioritizes the clean-up of production barriers, is also developing the next generation of hardware in the physical world model. Short-term engineering landings are perfectly compatible with long-term technological speculation and can coexist within the same subject.
The analyst ' s projections are essentially linear extrapolations of the established paradigm, which is good at estimating the hardware increments from mobile phones, servers, car-borne electronics: sample is adequate and variables are condensed, and the relevance of historical data supports predictions.
However, once the technology target is at the bottom of the equation, AI is no longer confined to the perception and interpretation of the information world, moving towards a proactive transformation of the physical structure, and the growth curve drawn from history sample is likely to lose its reference base.
Mask's judgement is not a standard industrial demand projection. It is more like a statement of objectives: to anchor the end point of the arrival of a smart person, and then to reverse the upper limit of the capacity required for the bottom infrastructure.
A set of narratives looked backwards, with the increments driven by historical patterns; another set to anchor the endpoint forward and push the limits of the hardware system.
For a long time to come, the storage industry will be in the middle of two sets of narratives, and no one will be sure which path is bound to take shape. Industry can only steadily move ahead with commercialization on visible steps, while immersing technology in the shadows for distant, or ultimately impossible, scenarios.
The core of the difference is never five years, 20 times or higher in level of demand, but rather we are willing to define the capacity boundary for silicon-based intelligence.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.