GPU load moves in milliseconds. A thermal machine moves in minutes.GPU 负荷以毫秒变化。热机以分钟变化。
That gap is six orders of magnitude, and it is the central engineering problem of on-site AI power. A training checkpoint drops tens of megawatts in seconds; a gas turbine needs a slower ramp than that, and a high-temperature fuel cell needs a far slower one or the stack cracks. ThermoPulse is the controller that sits between them.这个差距跨越六个数量级，也是现场 AI 供电的核心工程难题。一次训练检查点会在数秒内甩掉数十兆瓦；燃气轮机需要比这更缓的爬坡，而高温燃料电池需要缓得多——否则电堆会开裂。ThermoPulse 就是两者之间的那个控制器。
Four time domains that do not talk to each other.四个互不通话的时间域。
A conventional PLC or PID loop cannot bridge these. By the time a traditional controller has finished sampling, the stack has already taken the hit.常规 PLC 或 PID 回路无法跨越这些尺度。等传统控制器采样完成，电堆已经承受了冲击。
Compute load pulse算力负荷脉冲
Checkpoint saves and all-reduce synchronisation can swing a cluster by 30–70% of its draw almost instantaneously. This is the disturbance everything else has to absorb.检查点保存与 All-Reduce 同步可使集群取电在近乎瞬间波动 30%–70%。这就是其余环节必须吸收的扰动。
Solid-state power conversion固态电力变换
SiC and GaN switching in a solid-state transformer redirects power bidirectionally fast enough to meet the pulse head-on. This is the first line of defence, and it is the layer PowerHubX builds.固态变压器中的 SiC 与 GaN 开关能以足够快的速度双向调配功率，正面迎上脉冲。这是第一道防线，也是 PowerHubX 所构建的那一层。
Fast storage buffer快速储能缓冲
Supercapacitors, flywheels or a battery absorb and release across seconds, holding the line while the thermal machine begins to move. Sized against the residual, not the load.超级电容、飞轮或电池在数秒尺度上吸收与释放，在热机开始动作之前顶住。按残差定容，而非按负荷定容。
Turbine, engine and fuel cell燃机、发动机与燃料电池
These are heat engines. A turbine must avoid surge; a solid-oxide stack running at 700–900°C will crack under thermal stress if forced to follow a sharp load change. They need seconds to minutes of warning, and that warning is what the layers above exist to produce.这些都是热机。燃气轮机必须防喘振；工作在 700–900°C 的固体氧化物电堆若被迫跟随剧烈变负荷，会因热应力而开裂。它们需要数秒到数分钟的预警——而上面几层存在的意义，正是产生这个预警。
Three control layers, one coordinated response.三个控制层，一次协调响应。
The analogy is automotive. Enerfy does not build the engine or the battery — it builds the domain controller that makes them behave as one powertrain.这个类比来自汽车工业。Enerfy 不造发动机，也不造电池——我们造的是让它们表现为一套动力总成的域控制器。
Compute-aware prediction算力感知预测
Listen to the cluster scheduler, not just the power meter. A checkpoint that is about to be written is knowable before it draws current, and a physics-informed model turns that into a forward estimate of thermal drift and available margin.监听集群任务调度器，而不只是电表。一次即将写入的检查点，在它拉电流之前就是可知的；物理约束模型把这一信息转化为对热漂移与可用裕度的前瞻估计。
Multiphysics thermal coordination多物理场热工协调
Heat-exchanger flow and bypass valve positions coordinated against the predicted load, holding the stack within its thermal-stress envelope and the turbine clear of its surge line — while matching recovered heat to the cooling loop that wants it.依据预测负荷协调换热器流量与旁通阀开度，把电堆维持在热应力包络之内、把燃机保持在喘振线之外——同时将回收热量与需要它的冷却回路相匹配。
Semiconductor and storage execution半导体与储能执行
Solid-state conversion and fast storage absorb the transient in the moment it arrives, buying the layers above the seconds they need. Executed on the PowerHubX hardware layer, or on a third-party SST and battery system through the same interfaces.固态变换与快速储能在瞬变到来的当下将其吸收，为上面两层争取所需的数秒时间。可在 PowerHubX 硬件层上执行，也可通过相同接口在第三方固态变压器与储能系统上执行。
The machine you are protecting is the one you paid for.你要保护的那台机器，是你自己花钱买的。
On a grid-connected site, an unshaped load pulse is the utility’s problem. On a behind-the-meter site it is yours: it is thermal-stress cycles on a stack you own, surge margin on a turbine you own, and fuel burned at part load because the machine is being held back for safety it should not need.在并网站点上，一个未经整形的负荷脉冲是电力公司的问题。在表后站点上，它是你的问题：它意味着你自己电堆上的热应力循环、你自己燃机的喘振裕度，以及因为机组被按在低负荷「留安全余量」而多烧掉的燃料。
The gains therefore land in three places: equipment life, because the stack stops being cycled by every checkpoint; availability, because surge trips and thermal protection events stop taking the site down; and efficiency, because recovered heat can be matched to the cooling load instead of rejected, and the machine can run closer to its design point.因此收益体现在三处：设备寿命——电堆不再被每一次检查点反复循环；可用性——喘振跳机与热保护动作不再让站点停摆；能效——回收热量可以与冷却负荷匹配而非直接排放，机组也能更接近设计工况运行。
ThermoPulse is in development, with multiphysics co-simulation and controller-in-the-loop work underway alongside academic and industrial partners. It is not yet deployed on a commercial site. Performance figures discussed in technical conversations are modelled targets and are identified as such.ThermoPulse 正在开发中，多物理场联合仿真与控制器在环工作正与学术及工业伙伴同步推进。目前尚未在商业站点部署。技术沟通中涉及的性能数字均为模型目标值，并会如实注明。
One layer of three.三层中的一层。
Equipment layer装备层
Solid-state transformers and distributed power modules — the hardware that executes at millisecond scale.固态变压器与分布式功率模块——在毫秒尺度上执行的硬件。
Control layer控制层
The physical control brain across thermal, power, compute and cooling — seconds to minutes, with millisecond execution underneath.横跨热力、电力、算力与制冷的物理控制大脑——秒到分钟尺度，底层毫秒执行。
Reliability layer可靠性层
Dependency-aware risk and maintenance decisions — hours to months, read-only, above everything else.具备依赖关系认知的风险与检修决策——小时到月尺度，只读，位于全部之上。
Building on-site generation for an AI load? Let’s talk early.正在为 AI 负荷建设现场电源？越早沟通越好。
ThermoPulse changes what you should specify, so the most useful conversation happens before the generation and storage are ordered.ThermoPulse 会改变你应当如何选型，因此最有价值的沟通发生在发电与储能下单之前。