一天里连刷到三条视频。

一条说装光伏五年回本,阳光就是印钞机。一条说北方农村用太阳能取暖,一年省下几千块煤钱。一条说五千块攒台机器,大模型在家随便跑,从此不用再交订阅费。

三个行业,三个账号,一句一模一样的台词:花小钱,省大钱。

我把这三笔账各算了一遍。结论并不复杂:他们没撒谎,只是算术停在硬件报价那一行就收了笔。 往下还有三笔——电费和燃料、时间、机会成本,一笔没算。

技术上可行,和你该掏钱去做,是两件事。这两件事在一些内容里被合并成了一件,而这个合并是故意的。

一、光伏:分水岭不是年用电量,是你家白天有没有人

先把价格纠正一下。

组件价格确实崩了。2026 年 7 到 10 月的行业行情,TOPCon 组件每瓦 0.67 到 0.76 元,这个数字是真的,比 2023 年初的一块八跌了六成左右。但组件只是系统里的一件东西。同期分布式大 EPC 的均价是每瓦 2.70 到 2.78 元——那是几百千瓦级的工商业屋顶,设计、施工、并网摊得薄,才这个价。

户用五千瓦、十千瓦这种量级,屋顶分散、要爬梯子、要单独跑并网手续、要留渠道利润,单价只会更高,不会更低。

所以五千瓦的现实投入是一万五上下,十千瓦三万上下。你在视频里听到的「几千块装一套」,通常是把逆变器、支架、线缆、并网手续和人工全抹掉,只报组件那部分钱。组件占系统的四分之一,剩下的四分之三被藏起来了。

价格之外,真正的变量是发电量和你用掉多少。

华北地区的年等效满发小时数大概 1100 到 1300。五千瓦一年发 5500 到 6500 度,取中值 6000 度。这是你的原料。

然后是 2025 年的政策分界。2 月 9 日,国家发改委和能源局发布《关于深化新能源上网电价市场化改革促进新能源高质量发展的通知》,发改价格〔2025〕136 号。核心一句话:自 2025 年 6 月 1 日起,增量新能源项目的上网电量原则上全部进入电力市场,电价由市场交易形成。存量项目还有过渡保护,新项目没有。

行话翻译成人话:以前你发的电,电网按保障性价格全收;现在你自己去市场上卖。

而市场给光伏的电价,结构性地低。因为光伏的出力曲线和电价曲线是反的——中午发电最多的时候,正好是全系统电最不缺、价格最便宜的时候。山东电力现货市场 2023 年 5 月 1 日到 2 日的 48 小时里,有 32 个小时的实时电价是负的,最低到过 -85 元每兆瓦时。负电价的意思是:你发电,你还得给别人钱。

这不是山东的特殊情况,这是所有高比例光伏电网的物理宿命。你卖的那部分电,恰好是全系统最不值钱的那部分。

所以户用光伏的收益结构只剩一条:自用抵电价,余电贱卖。

居民一档电价五毛多,自用一度电省五毛多。余电上网按市场化电价结算,多数省份明显低于居民电价,还要被各类分摊再啃一口——具体数字看本省实施方案,别信安装商口头报的价。

按五千瓦、年发 6000 度、投入一万五算一笔账:

自用率自用省下的电费余电卖出收入年收益合计回收期
100%3360 元0 元3360 元4.5 年
70%2352 元450 元2802 元5.4 年
40%1344 元900 元2244 元6.7 年
20%672 元1200 元1872 元8.0 年

(自用电价按 0.56 元/度,余电按 0.25 元/度估算,实际以当地电价为准。)

这张表里唯一的变量是自用率。而自用率不由年用电量决定,它取决于你家的用电曲线和白天的出力曲线重不重合。

余电那一块越不值钱,这张表就越难看——这是 136 号文之后所有新项目的共同处境。

两个都是年用电 3000 度的家庭,一个是老人常年在家、白天开着空调和电视,一个是双职工晚上七点才进门——光伏在这两家的回收期能差出三年。前者自用率能到六成,后者可能只有两成,因为白天发的电,家里一个人都没有。

不带储能的户用系统,实际自用率通常落在 20% 到 40% 这一段。也就是回收期 6.7 到 8 年。视频里说的五年回本,要么按 100% 自用算(等于假设你白天把所有电都吃掉了),要么是拿 2021 年之前的价格和政策算的。

这话得说清楚:2021 年那套算法在当时是对的。组件一块八,有国家补贴,余电保障性全额收购按煤电基准价结算——那时候全额上网确实能算过来。是政策变了,不是那个时代的安装商都在骗人。变的是数学基础,没变的是话术。视频还在用旧公式,因为旧公式好卖。

想提自用率,得上储能。那是一笔额外的钱,还要接受电池衰减、循环寿命和十年级别的更换成本,逆变器本身也就八到十二年的寿命。这些在视频里全是空白。

最后是并网这道硬墙。并网容量、台区承载力、自然人还是非自然人备案、当地中午时段的消纳限制——不是你想卖就能卖,很多台区早就没有接入空间了。这一条短视频一个字都不提,因为提了就没法卖。

二、光热:省的是燃料钱,不是设备钱

这一块我愿意多说几句公道话:真空管集热器一套几千块,便宜是真的。

北方的农村自建房,冬天取暖烧煤烧柴,一年几千块的燃料支出也是真的。光热系统省下来的,就是这笔燃料钱。这事成立。

但纯光热在北方冬天物理上不成立。夜间没有太阳,连阴天没有太阳,而冷是最不讲道理的东西——它不会因为你的集热器白天很能干就晚上放你一马。

常规解法就两条,两条都得做:

一是加大保温水箱。 每平方米集热器配 50 到 75 升的储水容积,配上足够厚的保温层,可以把夜间的水温降控制在 5 到 8 度这个量级。水箱小了、保温差了,一夜掉二十度,第二天早上暖气片是凉的。

二是配辅助热源兜底。 电加热棒或者一台小锅炉,只在夜间和连阴天补温。这比全天烧锅炉省得多,也比纯光热可靠得多。

所以正确的形态是「光热为主 + 辅热兜底」。所有宣传「纯太阳能取暖、一分钱不花」的,不管画面拍得多暖和,都是不可信的——因为那不是设备不行,是物理不允许。

还有一笔账,短视频里也不会提:真正漏钱的地方不是设备,是房子。

北方农村自建房的墙体、窗户、屋面保温普遍很差,很多还是单层玻璃。热量从墙缝和玻璃跑掉的速度,比你用集热器收进来的速度快。房子不保温,集热面积要翻倍,水箱要加倍,辅热功率要加大——三笔钱同时涨,最后算下来比烧煤还贵。

还有一个很现实的约束:几百升到一两吨的水箱,你家有地方放吗?楼板承重够吗?这东西不是挂在墙上就完事的。

光热的门槛因此是三道:有没有辅热预算、房子保不保温、有没有地方放那个水箱。三道全过,它才是一笔划算的投入。

三、本地大模型:账算得对,但要算全

这一节我先站视频那边一半:五千块级别的洋垃圾——双路至强加一张 P40 或者改过显存的 2080Ti 22G——确实能跑 7B 到 14B 的量化模型。速度和可用性确实只是「能出字」。这个描述准确,比大部分博主诚实。

但硬件账只是第一笔。

电费。 推理时整机功率 250 到 350 瓦是轻载,满载跑大模型 400 瓦往上——P40 单卡 TDP 就 250 瓦,双路至强空载也不便宜。按一天 24 小时不关、平均 350 到 400 瓦算,一天 8 到 10 度电,一年 3000 到 3600 度,电费一千七到两千。这还没算夏天这台机器额外带给空调的负担。

折旧。 五千块的机器,三年摊下来一年一千六。三年后它还是那张 P40,但你已经不想用它跑新模型了。

时间。 驱动、CUDA 版本、量化格式、上下文长度、并发、显存溢出自动换 CPU——每一个都是坑,坑底都是你的周末。这些东西不产生任何产出,只是让你回到「能出字」的起点。

机会成本。 一个月两百块的编程订阅,一年两千四。五千块硬件加三年电费五千五,一共一万零五百,比三年订阅的七千二贵出三千多,而且你的时间还没算进去。三年之后,订阅用户用的是当年最好的模型,你用的是三年前的 P40。

这笔账算完,结论很清楚:自建不是省钱方案,是花钱买特定能力。

只有四类人值得自建:

数据不能出内网(医疗、法务、涉密项目);完全离线环境(野外、船、工厂内网);超高频低延迟调用(每秒几十次的自动化流水线,走 API 的延迟和费用都受不了);以及纯粹想学——这一类不算投资,算学费,学费花了就该花,别骗自己是在省钱。

「我想做个 app」「我想做个小游戏」属于短期 token 需求。两个月订阅比自建便宜一个数量级,而且你今天就能开始写代码,不用先花三天装驱动。

四、谁在卖

把三个领域摆在一起,利益链的结构几乎一模一样。

光伏内容,账号主体多是 EPC 商、组件经销商、区域代理。一套十千瓦户用系统的渠道毛利通常几千元,视频是获客漏斗的顶端,不是科普。

比卖设备更狠的是「屋顶租赁」和「合作共建」:户主不出现金,出的是屋顶和名字,背上十到十五年的贷款合同,发电收益先抵贷,抵完才是你的。这套模式出过大量纠纷——2023 年山东蓬莱,四十多户农户被「零投资建光伏电站」吸引签了合同,多数人贷款本金十万,最后演变成集体逾期。河南那边 2021 年就有媒体做过「别让农村光伏发电成套路」的调查。这类内容绝不会告诉你:余电上网的价格已经跌了。

本地部署内容,四层人在同一条链上:卖二手服务器的、卖显卡的(P40、2080Ti 22G、Mac Studio)、卖一体机的、卖教程课程的。转化路径是固定的三步——看起来硬核,建立技术信任,然后私域卖配置单或者整机。教程不是目的,教程是漏斗。

两边共用一套叙事技巧:只算硬件钱,不算电费、时间、机会成本;只讲技术可行,不讲经济可行。

技术可行是关于世界的判断,经济可行是关于你的判断。前者可以靠查资料得到,后者必须靠算自己的账。把前者包装成后者,是最省力也最有效的销售话术。

五、三个问题,问完再决定

项目过不去的信号结论
户用光伏白天家里没人;年用电不到 3000 度;主要指望卖电赚钱别装,回收期 8 年起
光热取暖不打算配辅热;房子没保温层;放不下大水箱别做纯光热,会冻着
本地大模型没有不可替代的场景;只是想试试;想省订阅费别买,订阅便宜一个数量级

我写过不少本地部署的文章,也帮人配过机器。那些文章告诉你的是技术上可行——什么卡能跑多大的模型,什么量化能塞进多少显存。

那不是让你为技术掏钱的理由。

光伏、光热、本机大模型,三个东西有一个共同的检验标准:它省下来的钱,能不能覆盖它额外带来的成本。 能覆盖的叫投资,不能覆盖的叫爱好。

爱好可以花钱。但你得知道自己在花钱买爱好,不是在买省钱。

三个门槛都过不去的话,剩下的就只有卖货的人赚钱。


数据边界:组件与 EPC 价格为 2026 年 7 至 10 月行业行情,136 号文原文为国家发改委、国家能源局发改价格〔2025〕136 号,负电价数据来自山东电力现货市场公开报道。发电量、自用率、回收期均为估算,各地电价、并网政策、屋顶条件差异很大,实际请以当地供电所和安装商的书面报价为准,不要采信口头承诺。

版权没有,随意转载。过意不去就加一句「原文发表于 cn-res.vip」,不加也无所谓。


Feasible to build, and worth paying for, are two different things.

Three videos, one afternoon. The first promised a solar array that pays for itself in five years. The second promised free winter heating from the sun. The third promised a five-thousand-yuan machine that runs large language models at home, so you never pay a subscription again.

Three industries, three accounts, one identical line: spend a little, save a lot.

I ran the numbers on all three. The conclusion isn't complicated. Nobody lied. They just stopped doing arithmetic the moment they reached the hardware price tag. Three more lines were left blank: electricity and fuel, time, opportunity cost.

That gap is not an oversight. It's the product.

I. Solar: the dividing line isn't annual consumption, it's whether anyone is home at noon

First, a correction on price.

Panel prices really did collapse. Through mid-to-late 2026, TOPCon modules trade at 0.67 to 0.76 yuan per watt — down more than sixty percent from roughly 1.8 yuan in early 2023. That part is true. But a module is one component. Distributed EPC pricing over the same period sits at 2.70 to 2.78 yuan per watt, and that's for commercial rooftops measured in hundreds of kilowatts, where design, labor, and grid connection get amortized thin.

A five-kilowatt residential job is scattered, requires ladders, needs its own interconnection paperwork, and has to carry a dealer's margin. The unit price goes up, never down.

Realistically: about 15,000 yuan for five kilowatts, about 30,000 for ten. The "a few thousand yuan for a whole system" you hear on video quietly deletes the inverter, the mounting rails, the wiring, the interconnection fees, and the labor. Modules are a quarter of a system. The other three quarters are hidden.

Beyond price, the real variables are how much you generate and how much of it you actually consume.

In northern China, equivalent full-load hours run about 1,100 to 1,300 per year. Five kilowatts yields 5,500 to 6,500 kWh; call it 6,000. That's your raw material.

Then came the policy break in 2025. On February 9, the National Development and Reform Commission and the National Energy Administration issued Document No. 136 — the notice on deepening market-based reform of new energy tariffs. The core sentence: from June 1, 2025, electricity from incremental new-energy projects enters the power market, and prices are set by trading. Legacy projects keep a transitional guarantee. New ones don't.

In plain language: the grid used to buy everything you produced at a guaranteed price. Now you sell into a market.

And the market prices solar power structurally low, because solar output is inversely correlated with price. Peak generation happens at noon, exactly when the system has the most surplus and the cheapest power. In China's Shandong spot market, over the 48 hours of May 1-2, 2023, real-time prices went negative for 32 hours, bottoming at -85 yuan per MWh. Negative pricing means you pay someone to take your electricity.

This isn't a Shandong quirk. It's the physical fate of any grid with a high solar share. The electricity you sell is precisely the electricity nobody wants.

So residential solar now has exactly one viable structure: offset your own consumption, dump the surplus cheap.

Retail residential power runs a bit above 0.5 yuan per kWh. Every kilowatt-hour you consume behind the meter saves that full amount. Every kilowatt-hour exported earns a market rate — usually well below retail, minus various levies. Check your province's implementation rules. Do not trust a verbal number from an installer.

Run it for five kilowatts, 6,000 kWh a year, 15,000 yuan installed:

Self-consumptionBill offsetExport incomeAnnual returnPayback
100%3,36003,3604.5 yrs
70%2,3524502,8025.4 yrs
40%1,3449002,2446.7 yrs
20%6721,2001,8728.0 yrs

(Retail at 0.56 yuan/kWh, export at 0.25 yuan/kWh, both estimates.)

The only meaningful variable is self-consumption rate. And that rate is not determined by how much electricity you use per year. It's determined by how well your household load curve overlaps with the solar output curve.

Two families can both use 3,000 kWh a year. One has a retiree at home running the air conditioner all afternoon. The other is a dual-income household where nobody walks in before seven. The payback period in those two houses can differ by three years. The first might hit sixty percent self-consumption. The second might hit twenty, because at noon there is literally nobody there to use the power being generated.

Without a battery, residential self-consumption typically lands between 20 and 40 percent. That's a payback of 6.7 to 8 years. The "five-year payback" claim on video either assumes you consume every kilowatt-hour at the exact moment it's produced, or it's using 2021 prices and 2021 policy.

To be fair: in 2021, that math worked. Modules at 1.8 yuan, a national subsidy, and guaranteed full purchase of surplus at the coal benchmark price. Full export genuinely penciled out. Policy changed; the installers of that era weren't all lying. The math changed while the script stayed. The script is still running because the old formula sells.

Raising your self-consumption rate means adding storage. That's another pile of money, plus battery degradation, cycle life, and replacement within a decade. The inverter itself only lasts eight to twelve years. None of this appears in the videos.

Then there's the hard wall: grid connection. Available capacity, transformer limits, whether you file as an individual or a corporate entity, and midday export restrictions in many areas. It is not simply a matter of deciding to sell. Plenty of distribution transformers have no headroom left at all. Not one video mentions this, because mentioning it ends the sale.

II. Solar thermal: it saves fuel, not equipment

Credit where due: evacuated-tube collectors cost a few thousand yuan a set. That cheapness is real.

Rural houses in the north burning coal or wood through the winter spend a few thousand a year on fuel. That's real too. A solar thermal system does capture that fuel cost. The premise holds.

But pure solar thermal does not work in a northern winter. There is no sun at night. There is no sun for days on end. And cold does not negotiate — it will not go easy on you because your collectors had a productive afternoon.

There are exactly two fixes, and you need both.

One: a bigger, better-insulated tank. Fifty to seventy-five liters of storage per square meter of collector area, with serious insulation, keeps overnight temperature drop in the five-to-eight-degree range. Undersize the tank or skimp on insulation and you'll lose twenty degrees overnight. The radiators will be cold by morning.

Two: a backup heat source. An electric element or a small boiler, kicking in only at night and during overcast stretches. Far cheaper than running a boiler around the clock, and far more reliable than thermal alone.

The correct configuration is solar thermal as primary, backup as insurance. Anyone advertising "pure solar heating, zero cost" is selling something physics doesn't support — not because the equipment is bad, but because the physics refuses.

There's another cost the videos never mention: the expensive part isn't the equipment, it's the house.

Rural self-built homes in the north have poor wall, window, and roof insulation. Many still have single glazing. Heat escapes through gaps and glass faster than your collectors can bring it in. Without insulation, you'll need double the collector area, double the tank, and more backup capacity — three costs rising at once, until the system costs more than just burning coal.

One more practical constraint: a tank holding several hundred liters, or a couple of tons, needs floor space and load-bearing capacity. It doesn't hang on a wall.

So solar thermal has three gates: a backup heat budget, an insulated envelope, and somewhere to put the tank. Clear all three and it's a sound investment.

III. Local models: the hardware math is honest, the rest isn't included

I'll grant the videos half a point here. A five-thousand-yuan pile of decommissioned server hardware — dual Xeons plus a P40 or a modified 22 GB 2080 Ti — really will run 7B to 14B quantized models. And yes, the speed and quality are best described as "it produces text." That description is accurate, and more honest than most influencers manage.

But hardware is only the first line.

Electricity. 250 to 350 watts is light load. Under real inference a dual-Xeon box runs 400 watts and up — a single P40 has a 250-watt TDP, and idle dual Xeons aren't cheap either. At 350 to 400 watts averaged over 24 hours, that's 8 to 10 kWh a day, 3,000 to 3,600 kWh a year, roughly 1,700 to 2,000 yuan. That's before counting what the machine adds to your air conditioning bill in summer.

Depreciation. Five thousand yuan over three years is 1,600 a year. In three years it's still a P40, and you no longer want to run current models on it.

Time. Drivers, CUDA versions, quantization formats, context length, concurrency, automatic CPU offload when VRAM overflows. Every one is a rabbit hole, and every hole is a weekend. None of it produces anything. It just gets you back to the starting line of "it produces text."

Opportunity cost. A coding subscription runs about thirty dollars a month, 2,400 yuan a year. The hardware plus three years of electricity costs roughly 10,500 yuan — more than three years of subscription at 7,200, with your time still unpriced. And after those three years, the subscriber is using the best model available that year while you're using a three-year-old P40.

Run the whole thing and the conclusion is unambiguous. Self-hosting isn't a cost-saving measure. It's a purchase of a specific capability.

Only four categories justify it. Data that legally cannot leave your network — medical, legal, classified. Fully offline environments — field sites, ships, air-gapped factory floors. Extremely high-frequency, low-latency calls — pipelines making dozens of requests per second, where API latency and billing both break down. And people who simply want to learn. That last one isn't an investment, it's tuition. Pay it gladly, but don't tell yourself you're saving money.

"I want to build an app" or "I want to make a small game" is short-term token demand. Two months of subscription costs an order of magnitude less than building, and you can start writing code today instead of spending three days on drivers.

IV. Who's selling

Put the three fields side by side and the structure is nearly identical.

Solar content is mostly run by EPC contractors, module distributors, and regional agents. Channel margin on a ten-kilowatt residential job is typically a few thousand yuan. The video is the top of a lead funnel, not public education.

Nastier than selling equipment is the rooftop lease: the homeowner puts up no cash, but contributes the roof and their name, signing a ten-to-fifteen-year loan, with generation revenue servicing the debt first. This model has generated a long trail of disputes. In 2023 in Penglai, Shandong, more than forty farming households signed up for "zero-investment solar," most carrying loan principals around 100,000 yuan, and ended up in collective default. Henan media were already documenting the pattern back in 2021. None of this content will tell you that surplus export prices have collapsed.

Local deployment content has four layers on one chain: used server sellers, GPU sellers (P40, 22 GB 2080 Ti, Mac Studio), prebuilt-machine vendors, and course sellers. The conversion path is fixed — look hardcore, build technical trust, sell a configuration sheet or a full machine in a private chat. The tutorial isn't the product. The tutorial is the funnel.

Both sides share one narrative trick: count the hardware, skip the electricity, the time, and the opportunity cost. Talk about technical feasibility, never economic feasibility.

Technical feasibility is a claim about the world. Economic feasibility is a claim about you. The first can be researched. The second can only be settled by running your own numbers. Packaging the first as the second is the cheapest and most effective sales technique ever invented.

V. Three questions to answer before you spend anything

PurchaseDisqualifying signalVerdict
Residential solarNobody home during the day; under 3,000 kWh a year; counting on export incomeDon't. Payback is 8 years plus.
Solar thermal heatingNo backup heat budget; uninsulated house; no room for a large tankDon't do pure thermal. You'll freeze.
Local LLM rigNo irreplaceable use case; just curious; trying to dodge subscription feesDon't. Subscription is an order of magnitude cheaper.

I've written plenty of local deployment guides, and I've spec'd machines for people. Those guides tell you what's technically possible — which card runs which model size, which quantization fits in how much VRAM.

That was never a reason to spend money on the technology.

Solar, solar thermal, local models — one test applies to all three. Does the money it saves cover the costs it adds? If yes, it's an investment. If no, it's a hobby.

Hobbies are fine. Just know that you're buying a hobby, not buying savings.

Fail all three gates, and the only person making money is the one selling.


Data notes: module and EPC prices reflect mid-to-late 2026 market conditions. Document No. 136 is NDRC/NEA Fagai Jiage [2025] No. 136. Negative price data comes from public reporting on the Shandong spot market. Generation, self-consumption, and payback figures are estimates; local tariffs, interconnection rules, and roof conditions vary widely. Get a written quote from your local utility and installer, and treat verbal promises as worthless.

No copyright reserved — share it freely. If you feel like it, a credit line to cn-res.vip is appreciated. Or don't. Either way.