我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
GPU三年归零不是会计魔术
视频
音频
原始脚本
GPU 三年实质归零,不是会计魔术,是硅谷 AI 行业藏不住的生死宿命。 在 AI 算力席卷全球的当下,外界普遍对 GPU 有着根深蒂固的认知误区。 一张显卡,只要硬件没损坏,能点亮运行,就始终具备价值。 所谓三年折旧归零,不过是企业做账的财务手段,是虚头巴脑的账面游戏。 但事实上, GPU 三年实质归零,从来都不是会计层面的数字操作,而是商业价值、物理法则、行业规则三重因素共同作用的必然结果。 是整个硅谷 AI 算力行业心照不宣却又极力掩盖的核心真相。 那个流传在算力行业的极端冷笑话,早已道尽了背后的残酷逻辑。 而马斯克解散 xai 算力集群,与 anthropic 达成长期租约的操作,更是精准踩中了这一行业命脉,撕开了云巨头们的财务遮羞布。 一、行业冷笑话。 未开封的 GPU 反而比满负荷空转更值钱。 国内民营算力行业曾出现过一个令人唏嘘的真实案例。 堪称 GPU 行业的黑色幽默。 有民营资本斥巨资自建算力中心,大批量采购高端英伟达 GPU 原本计划对外出租盈利。 却因市场遇冷,客户流失,最终陷入无单可接的困境。 最终出现了反常识的结局,一部分 GPU 被顺利装机上架。 二十四小时不间断通电运行,却始终没有租赁业务,只能空转耗电。 另一部分 GPU 因来不及部署,一直原封不动放在仓库。 从未拆封,从未通电。 三年后复盘,未拆封的 GPU 反而比空转三年的 GPU 亏损更少,残值更高。 满负荷空转的 GPU 不仅每天产生巨额电费成本,持续的高负载、高温环境让芯片快速老化,即便外观完好,也已是重度损耗的工业旧卡。 几乎无人问津。 而仓库里未拆封的 GPU 没有任何电力消耗,半导体元件处于完全休眠状态,好歹保留了全新品相。 残值略高一些。 但这只是残酷现实里的微小差别,即便全新未拆封,三年之后也基本卖不出去。 正如英伟达创始人黄仁勋直言不讳。 如今的 H100显卡送都送不出去,这不是硬件损坏,而是 GPU 的核心商业价值已经被彻底清零。 二、商业价值死刑。 英伟达激进迭代,旧卡彻底失去训练市场 GPU 的核心价值。 从来不是能运行,而是大模型训练的效率与性价比。 这也是头部 AI 模型公司唯一看重的指标。 英伟达的迭代节奏早已打破常规电子产品的更新周期,从最初两年一代。 压缩至如今,12~18个月一轮大架构升级,H100、 H200、 Blackwall 等系列轮番登场。 每一代新卡都实现性能翻倍,能耗降低30%~50%,互联带宽、 MO 架构适配性全面碾压上代产品。 对于 Anthropic OpenAI 这类头部模型公司而言,时间成本远高于硬件租赁成本。 其核心算法团队人均年薪百万美元起步。 顶尖技术人才薪资更是达到千万甚至上亿级别,每一天的人力成本、研发成本都是天文数字。 模型早一天训练完成,早一天上线商业化。 就能抢占市场先机,带来的收益远超租金差价。 他们对 GPU 的需求,如同顶级电竞玩家对硬件的追求,只认最新、最快、最强的产品。 不计较短期成本差异。 旧卡即便租金更低,训练速度却慢二到三倍,不仅拉长研发周期,还会错失市场窗口,完全得不偿失。 这就导致高端训练市场对三年以上的旧卡直接实行零准入。 旧卡哪怕全新未拆封,在模型训练领域也毫无用武之地。 最核心的商业价值彻底归零,这是市场用脚投票的结果,绝非会计数字可以改变。 三、物理法则绝杀,高温、高负载。 半导体三年实质衰变报废,很多人疑惑 GPU 放在那里好好的,没磕没碰,为什么三年就不能用了?答案藏在半导体物理的铁律里,也藏在数据中心真实的运行环境中。 个人使用的 GPU 日常负载仅30%左右,温度控制在五六十度,偶尔使用,用五六年依旧能正常运行。 带数据中心的 AI 训练 GPU 常年维持百分之八十到百分之一百满负荷运转,核心温度长期保持在七八十度以上,二十四小时全年无休。 在持续高温、高电压、高负载的环境下, GPU 内部的硅晶片会发生电子迁移、晶格损伤、晶体管隐性击穿。 这种损耗不是瞬间损坏,而是可靠性逐年断崖式下跌。 第一年稳定性拉满,故障率极低。 第二年隐性故障开始频发。 第三年起,随机报错、节点掉线、算力抖动的概率大幅飙升,完全无法满足大模型训练的稳定性要求。 大模型训练采用数千张 GPU 集群并行作业。 哪怕一两张卡出现故障,整个训练任务就会彻底崩盘。 前期投入的巨额算力成本、时间成本全部付诸东流,必须推倒重来。 这是模型公司最无法承受的噩梦。 因此,他们绝不会使用服役超过3年的旧卡参与核心训练。 不是卡不能运行,而是赌不起可靠性风险。 从物理损耗层面来看,三年期的 GPU 即便外观完好,也已失去工业级大规模使用的价值,实质进入报废阶段。 四、最后的遮羞布。 旧卡只能做推理,却撑不起资产原值。 有人会提出质疑,旧卡不能做训练,还能做推理业务,怎么能算归零?这恰恰是云巨头们用来掩盖财务漏洞的最后遮羞布。 但根本无法改变 GPU 三年贬值的本质。 英伟达高端 GPU 从设计之初就是为大模型训练量身打造架构、显存。 互联方案全部偏向训练场景,用来做推理属于典型的大马拉小车,性能严重过剩,能效比极差,成本控制完全不占优势。 云巨头们只能将三年到期的旧卡下沉到低端推理边缘业务,低优先级请求队列,进行内部消化。 这类业务租金极低,毛利微薄。 只能勉强覆盖电费和运维成本,完全无法回收当初的巨额采购成本。 旧卡即便能产生微弱现金流,也只是残值兜底,其资产价值早已远低于原始采购价。 从商业投资角度来看,就是实质归零。 五、硅谷财务黑洞。 巨头的5~6年折旧只是自欺欺人的会计魔术,面对 GPU 三年实质归零的行业现实。 硅谷云巨头们却集体选择了财务操作。 谷歌、微软将 GPU 折旧年限定为6年, Meta 定为5.5年,甲骨文同样采用6年折旧。 仅有亚马逊相对谨慎,将训练集群折旧年限定为5年,还在2025年主动计提了资产减值。 按照真实3年折旧与账面5~6年折旧计算,同等价值的 GPU 集群,巨头们每年能虚增近一半的利润。 大空头迈克尔伯里测算,2026~2028年,硅谷五大云巨头仅通过延长 GPU 折旧,就能累计虚增利润1760亿美元。 Meta 甲骨文利润高估比例甚至超过20%。 他们刻意隐瞒 GPU 真实寿命,靠推理业务消化旧卡残值,假装资产长期保值。 将巨额隐形亏损向后拖延。 但这种操作治标不治本。 2027~2028年,首批大规模部署的 H100集群将满3年。 届时旧卡彻底失去利用价值,巨头们必将面临千亿级资产减值,AI 算力行业的财务泡沫终将破裂。 六,马斯克的清醒。 快建快租,用三年真实折旧,跳出行业陷阱。 反观马斯克,解散 xai 算力集群,将22万张 GPU 长期出租给 anthropic。 恰恰是看透了 GPU 三年归零的残酷现实,做出的最优决策。 马斯克没有云生态加持,无法像谷歌、亚马逊那样玩会计魔术。 若将 GPU 握在自己手中做模型训练,不仅要承担巨额电费、运维成本,还会面临三年后资产彻底贬值的风险,最终沦为一堆废铁。 因此,他选择短平快自建算力,用天然气涡轮机解决电力审批难题。 快速建成集群后,直接与 Anthropic 签订长期租约。 每年锁定50亿美元稳定现金流,严格按照三年真实寿命计提折旧,将 GPU 迭代贬值、物理损耗的风险全部转移给承租方。 三年折旧期满后,后续租金全部变为纯利润。 总结, GPU 三年实质归零,从来都不是会计层面的虚数,而是商业规律、物理法则共同铸就的行业宿命。
修正脚本
GPU 三年实质归零,不是会计魔术,是硅谷 AI 行业藏不住的生死宿命。 在 AI 算力席卷全球的当下,外界普遍对 GPU 有着根深蒂固的认知误区。 一张显卡,只要硬件没损坏,能点亮运行,就始终具备价值。 所谓三年折旧归零,不过是企业做账的财务手段,是虚头巴脑的账面游戏。 但事实上, GPU 三年实质归零,从来都不是会计层面的数字操作,而是商业价值、物理法则、行业规则三重因素共同作用的必然结果。 是整个硅谷 AI 算力行业心照不宣却又极力掩盖的核心真相。 那个流传在算力行业的极端冷笑话,早已道尽了背后的残酷逻辑。 而马斯克解散 xai 算力集群,与 anthropic 达成长期租约的操作,更是精准踩中了这一行业命脉,撕开了云巨头们的财务遮羞布。 一、行业冷笑话。 未开封的 GPU 反而比满负荷空转更值钱。 国内民营算力行业曾出现过一个令人唏嘘的真实案例。 堪称 GPU 行业的黑色幽默。 有民营资本斥巨资自建算力中心,大批量采购高端英伟达 GPU,原本计划对外出租盈利。 却因市场遇冷,客户流失,最终陷入无单可接的困境。 最终出现了反常识的结局,一部分 GPU 被顺利装机上架。 二十四小时不间断通电运行,却始终没有租赁业务,只能空转耗电。 另一部分 GPU 因来不及部署,一直原封不动放在仓库。 从未拆封,从未通电。 三年后复盘,未拆封的 GPU 反而比空转三年的 GPU 亏损更少,残值更高。 满负荷空转的 GPU 不仅每天产生巨额电费成本,持续的高负载、高温环境让芯片快速老化,即便外观完好,也已是重度损耗的工业旧卡。 几乎无人问津。 而仓库里未拆封的 GPU 没有任何电力消耗,半导体元件处于完全休眠状态,好歹保留了全新品相。 残值略高一些。 但这只是残酷现实里的微小差别,即便全新未拆封,三年之后也基本卖不出去。 正如英伟达创始人黄仁勋直言不讳。 如今的 H100显卡送都送不出去,这不是硬件损坏,而是 GPU 的核心商业价值已经被彻底清零。 二、商业价值死刑。 英伟达激进迭代,旧卡彻底失去训练市场。GPU 的核心价值,从来不是能运行,而是大模型训练的效率与性价比。 这也是头部 AI 模型公司唯一看重的指标。 英伟达的迭代节奏早已打破常规电子产品的更新周期,从最初两年一代。 压缩至如今12~18个月一轮大架构升级,H100、 H200、 Blackwell 等系列轮番登场。 每一代新卡都实现性能翻倍,能耗降低30%~50%,互联带宽、 MO 架构适配性全面碾压上代产品。 对于 Anthropic OpenAI 这类头部模型公司而言,时间成本远高于硬件租赁成本。 其核心算法团队人均年薪百万美元起步。 顶尖技术人才薪资更是达到千万甚至上亿级别,每一天的人力成本、研发成本都是天文数字。 模型早一天训练完成,早一天上线商业化。 就能抢占市场先机,带来的收益远超租金差价。 他们对 GPU 的需求,如同顶级电竞玩家对硬件的追求,只认最新、最快、最强的产品。 不计较短期成本差异。 旧卡即便租金更低,训练速度却慢二到三倍,不仅拉长研发周期,还会错失市场窗口,完全得不偿失。 这就导致高端训练市场对三年以上的旧卡直接实行零准入。 旧卡哪怕全新未拆封,在模型训练领域也毫无用武之地。 最核心的商业价值彻底归零,这是市场用脚投票的结果,绝非会计数字可以改变。 三、物理法则绝杀,高温、高负载。 半导体三年实质衰变报废,很多人疑惑 GPU 放在那里好好的,没磕没碰,为什么三年就不能用了?答案藏在半导体物理的铁律里,也藏在数据中心真实的运行环境中。 个人使用的 GPU 日常负载仅30%左右,温度控制在五六十度,偶尔使用,用五六年依旧能正常运行。 但数据中心的 AI 训练 GPU 常年维持百分之八十到百分之一百满负荷运转,核心温度长期保持在七八十度以上,二十四小时全年无休。 在持续高温、高电压、高负载的环境下, GPU 内部的硅晶片会发生电子迁移、晶格损伤、晶体管隐性击穿。 这种损耗不是瞬间损坏,而是可靠性逐年断崖式下跌。 第一年稳定性拉满,故障率极低。 第二年隐性故障开始频发。 第三年起,随机报错、节点掉线、算力抖动的概率大幅飙升,完全无法满足大模型训练的稳定性要求。 大模型训练采用数千张 GPU 集群并行作业。 哪怕一两张卡出现故障,整个训练任务就会彻底崩盘。 前期投入的巨额算力成本、时间成本全部付诸东流,必须推倒重来。 这是模型公司最无法承受的噩梦。 因此,他们绝不会使用服役超过3年的旧卡参与核心训练。 不是卡不能运行,而是赌不起可靠性风险。 从物理损耗层面来看,三年期的 GPU 即便外观完好,也已失去工业级大规模使用的价值,实质进入报废阶段。 四、最后的遮羞布。 旧卡只能做推理,却撑不起资产原值。 有人会提出质疑,旧卡不能做训练,还能做推理业务,怎么能算归零?这恰恰是云巨头们用来掩盖财务漏洞的最后遮羞布。 但根本无法改变 GPU 三年贬值的本质。 英伟达高端 GPU 从设计之初就是为大模型训练量身打造架构、显存。 互联方案全部偏向训练场景,用来做推理属于典型的大马拉小车,性能严重过剩,能效比极差,成本控制完全不占优势。 云巨头们只能将三年到期的旧卡下沉到低端推理边缘业务,低优先级请求队列,进行内部消化。 这类业务租金极低,毛利微薄。 只能勉强覆盖电费和运维成本,完全无法回收当初的巨额采购成本。 旧卡即便能产生微弱现金流,也只是残值兜底,其资产价值早已远低于原始采购价。 从商业投资角度来看,就是实质归零。 五、硅谷财务黑洞。 巨头的5~6年折旧只是自欺欺人的会计魔术,面对 GPU 三年实质归零的行业现实。 硅谷云巨头们却集体选择了财务操作。 谷歌、微软将 GPU 折旧年限定为6年, Meta 定为5.5年,甲骨文同样采用6年折旧。 仅有亚马逊相对谨慎,将训练集群折旧年限定为5年,还在2025年主动计提了资产减值。 按照真实3年折旧与账面5~6年折旧计算,同等价值的 GPU 集群,巨头们每年能虚增近一半的利润。 大空头迈克尔伯里测算,2026~2028年,硅谷五大云巨头仅通过延长 GPU 折旧,就能累计虚增利润1760亿美元。 Meta 甲骨文利润高估比例甚至超过20%。 他们刻意隐瞒 GPU 真实寿命,靠推理业务消化旧卡残值,假装资产长期保值。 将巨额隐形亏损向后拖延。 但这种操作治标不治本。 2027~2028年,首批大规模部署的 H100集群将满3年。 届时旧卡彻底失去利用价值,巨头们必将面临千亿级资产减值,AI 算力行业的财务泡沫终将破裂。 六、马斯克的清醒。 快建快租,用三年真实折旧,跳出行业陷阱。 反观马斯克,解散 xai 算力集群,将22万张 GPU 长期出租给 anthropic。 恰恰是看透了 GPU 三年归零的残酷现实,做出的最优决策。 马斯克没有云生态加持,无法像谷歌、亚马逊那样玩会计魔术。 若将 GPU 握在自己手中做模型训练,不仅要承担巨额电费、运维成本,还会面临三年后资产彻底贬值的风险,最终沦为一堆废铁。 因此,他选择短平快自建算力,用天然气涡轮机解决电力审批难题。 快速建成集群后,直接与 Anthropic 签订长期租约。 每年锁定50亿美元稳定现金流,严格按照三年真实寿命计提折旧,将 GPU 迭代贬值、物理损耗的风险全部转移给承租方。 三年折旧期满后,后续租金全部变为纯利润。 总结, GPU 三年实质归零,从来都不是会计层面的虚数,而是商业规律、物理法则共同铸就的行业宿命。
英文翻译
GPU Reaches Zero Actual Value in Three Years: Not an Accounting Trick, but the Inescapable Fate of Silicon Valley's AI Industry. Amid the global AI computing frenzy, the public holds a deeply ingrained misconception about GPUs. A graphics card, as long as its hardware is intact and it can power on and run, is believed to always retain value. The so-called three-year depreciation to zero is merely a financial accounting maneuver, a paper game. But in reality, the GPU's actual value dropping to zero in three years has never been a numerical accounting exercise—it is the inevitable outcome of three interlocking factors: commercial value, physical laws, and industry rules. It is the core truth that the entire Silicon Valley AI computing industry tacitly acknowledges yet desperately conceals. The dark joke circulating in the computing industry has long laid bare the brutal logic behind it. And Musk's move to disband xAI's computing cluster and sign a long-term lease with Anthropic precisely hits this industry nerve, tearing off the financial fig leaf of the cloud giants. **1. The Industry Joke: Unopened GPUs Are Worth More Than Those Running Idle at Full Load** A real-life case in China's private computing industry is a tragic illustration. It is the black humor of the GPU sector. A private capital firm poured huge sums into building its own computing center, purchasing large quantities of high-end NVIDIA GPUs, originally planning to rent them out for profit. But due to a cold market and lost customers, they ended up with no orders. The outcome defied common sense: some GPUs were successfully installed and racked. They ran 24/7 with power, yet never had any leasing business—just idled, consuming electricity. Another batch of GPUs, not yet deployed, sat untouched in the warehouse. Never opened, never powered on. Three years later, a review showed that the unopened GPUs actually incurred less loss and had higher residual value than those that had idled for three years. The GPUs that ran idle at full load not only generated massive electricity costs daily, but the continuous high load and high temperature rapidly aged the chips. Even if they looked intact, they were heavily worn industrial old cards. Almost no one wanted them. Meanwhile, the unopened GPUs in the warehouse consumed no electricity. The semiconductor components were in a completely dormant state, at least retaining a brand-new appearance. Their residual value was slightly higher. But this is just a tiny difference in a cruel reality—even brand new and unopened, after three years they are basically unsellable. As NVIDIA founder Jensen Huang bluntly stated: "Today, H100s can't even be given away." It's not hardware damage, but the core commercial value of the GPU has been completely wiped out. **2. The Death Sentence of Commercial Value: NVIDIA's Aggressive Iteration Renders Old Cards Useless for Training** The core value of a GPU has never been about whether it can run, but about the efficiency and cost-effectiveness of large model training. That is the only metric top AI model companies care about. NVIDIA's iteration pace has long broken the normal update cycle of consumer electronics, shrinking from a two-year generation to a major architecture upgrade every 12–18 months. The H100, H200, Blackwell, and other series roll out one after another. Each new generation doubles performance, reduces power consumption by 30%–50%, and completely outperforms the previous generation in interconnect bandwidth and MoE architecture compatibility. For top model companies like Anthropic and OpenAI, time cost far outweighs hardware rental cost. Their core algorithm teams earn at least a million dollars per person annually. Top technical talent can command salaries in the tens or even hundreds of millions. Every day of labor and R&D costs is astronomical. If a model can be trained and commercialized one day earlier, it can seize market advantage, generating far more revenue than any rental price difference. Their demand for GPUs is like top esports players chasing the newest, fastest, and strongest hardware. They don't care about short-term cost differences. Even if an older card has a lower rental price, its training speed is two to three times slower, not only extending the R&D cycle but also missing market windows—entirely counterproductive. This leads to a zero-tolerance policy for cards older than three years in the high-end training market. Even if an old card is brand new and unopened, it has no use in model training. Its core commercial value is completely zero—a market vote that no accounting number can change. **3. The Absolute Rule of Physics: High Temperature and High Load Cause Semiconductor Degradation and Failure in Three Years** Many wonder: a GPU sits there perfectly fine, no bumps or bruises—why can't it be used after three years? The answer lies in the iron law of semiconductor physics and the real operating environment of data centers. A personal GPU typically runs at around 30% load, with temperatures controlled at 50–60°C. Used occasionally, it can still function normally after five or six years. But an AI training GPU in a data center runs at 80%–100% full load year-round, with core temperatures consistently above 70–80°C, 24/7. Under sustained high temperature, high voltage, and high load, the silicon die inside the GPU suffers electromigration, lattice damage, and hidden transistor breakdown. This degradation isn't sudden failure; it's a cliff-like drop in reliability year by year. Year one: maximum stability, extremely low failure rate. Year two: latent faults begin to appear frequently. By year three, the probability of random errors, node dropouts, and computing fluctuations spikes dramatically, making it completely unable to meet the stability requirements of large model training. Large model training uses clusters of thousands of GPUs working in parallel. Even if one or two cards fail, the entire training task collapses. All the massive upfront computing cost and time are wasted, forcing a restart from scratch. This is the worst nightmare for model companies. Thus, they absolutely will not use cards older than three years for core training. It's not that the cards can't run—it's that they can't afford the reliability risk. From a physical wear perspective, a three-year-old GPU, even if it looks intact, has lost its value for industrial-scale use and has effectively entered the scrap stage. **4. The Last Fig Leaf: Old Cards Can Only Do Inference, But Can't Support Their Original Asset Value** Some might argue: old cards can't do training, but they can still do inference—how can that be zero? This is precisely the last fig leaf cloud giants use to cover their financial holes. But it cannot change the essence of GPU depreciation over three years. NVIDIA's high-end GPUs were designed from the ground up for large model training—architecture, memory, and interconnect all favor training scenarios. Using them for inference is like a large horse pulling a small cart: massive performance surplus, terrible energy efficiency, and no cost advantage. Cloud giants can only downgrade three-year-old cards to low-end inference, edge services, and low-priority request queues for internal consumption. Such services generate very low rental income and thin margins. They barely cover electricity and maintenance costs, let alone recover the huge initial procurement costs. Even if old cards produce a trickle of cash flow, it's just salvaging residual value. Their asset value is already far below the original purchase price. From a commercial investment perspective, that is zero in substance. **5. Silicon Valley's Financial Black Hole: The 5–6 Year Depreciation of Giants Is Just Self-Deceiving Accounting Magic** Facing the industry reality of GPU value dropping to zero in three years, Silicon Valley's cloud giants have collectively chosen accounting maneuvers. Google and Microsoft set GPU depreciation at six years, Meta at 5.5 years, and Oracle also at six years. Only Amazon is relatively cautious, setting training cluster depreciation at five years and proactively recognizing asset impairment in 2025. Compared to the actual three-year depreciation versus the book depreciation of 5–6 years, for GPU clusters of equal value, the giants can inflate their profits by nearly half each year. According to Michael Burry's calculations, from 2026 to 2028, the top five Silicon Valley cloud giants could cumulatively inflate profits by $176 billion just by extending GPU depreciation. The overstatement of profits for Meta and Oracle could exceed 20%. They deliberately hide the true lifespan of GPUs, rely on inference business to salvage old card value, pretend assets retain value long-term, and push massive hidden losses into the future. But this approach is a temporary fix, not a cure. In 2027–2028, the first large-scale deployed H100 clusters will reach three years of age. By then, old cards will lose all utility, and the giants will face hundreds of billions in asset impairments. The financial bubble in the AI computing industry will finally burst. **6. Musk's Clarity: Build Fast, Rent Out, Use Three-Year Real Depreciation to Escape the Industry Trap** In contrast, Musk's decision to disband xAI's computing cluster and lease 220,000 GPUs long-term to Anthropic is precisely the optimal move after seeing through the brutal reality of GPU zeroing in three years. Musk doesn't have a cloud ecosystem to play the accounting tricks like Google or Amazon. If he held the GPUs himself for model training, he would not only bear huge electricity and maintenance costs but also face the risk of total asset depreciation after three years, turning them into scrap metal. So he chose a fast, low-cost approach: build the cluster quickly using natural gas turbines to bypass power approval issues. Then sign a long-term lease with Anthropic. Lock in a stable $5 billion annual cash flow, depreciate strictly according to the real three-year lifespan, and transfer all risks of GPU iteration depreciation and physical wear to the lessee. After the three-year depreciation period, subsequent rental income becomes pure profit. **Conclusion: GPU Reaching Zero Actual Value in Three Years Has Never Been an Accounting Fiction, but an Industry Destiny Forged by Commercial Laws and Physical Rules.**
back to top