401MW。
这是什么概念？一个标准中型核电站的装机容量大约在1000MW左右。7月6日，美国数字基础设施公司TeraWulf宣布，与Anthropic签署了一份为期20年的数据中心租赁协议——Anthropic将租用位于肯塔基州Hawesville的Justified Data园区，可容纳约401MW的关键IT负载，预计为TeraWulf创造190亿美元（约合1292亿元人民币）的全周期收入。这不是一份普通的云服务合同，这是AI产业有史以来最长的算力锁定期。
20年。这个时间尺度意味着什么？意味着这份合同到期的时候，今天刚上大学的学生已经年近不惑；意味着iPhone才发布了19年；意味着谷歌成立也才28年。一家成立不到5年的AI公司，为自己锁定了一份比大多数科技公司历史还长的基础设施合约。
算力合约的三级跳：从年付到20年
回顾AI产业的算力采购史，你会看到一条清晰的时间轴拉长曲线。
2023年到2024年，大模型公司主要按年签云服务合同。OpenAI和微软的协议最初是每年一议，大家对算力需求的预估还停留在"今年需要多少张H100"的层面。那时候采购GPU像是在菜市场买菜——今年买1万张，明年看情况再加。
2025年，合约期拉长到3-5年。OpenAI与微软签署了多年期超级计算机协议，Google向Anthropic承诺未来5年交付5GW（5000MW）算力，从2027年开始陆续上线。5年是什么概念？是一个数据中心从规划到建成满负荷运转的周期。这时候AI公司开始意识到：算力不是随用随买的水电，而是需要提前数年锁定的战略资源。
2026年7月，Anthropic把这个时间尺度拉到了20年。这已经不是科技公司的采购逻辑了——这是电力公司、铁路公司、自来水公司的逻辑。你见过哪家互联网公司为服务器机房签20年租约？没有。但你见过电力公司为一座水电站签20年电力采购协议吗？常见得很。
这份合约的细节也很说明问题。园区分阶段开发，预计2027年下半年初步投用，2028年初达到最大装机容量。也就是说Anthropic现在签约，要等一年多才能用上第一批电。与此同时，TeraWulf在同一公告中宣布，将其在得克萨斯州Abernathy合资项目（168MW）中持有的50.1%股权全部出售给合作伙伴Fluidstack。这不是扩张，这是All in——卖掉其他项目，集中全部资源服务Anthropic这一个客户。
为什么是20年？三个残酷现实
Anthropic不是慈善家，TeraWulf也不是。一份20年、190亿美元的合约能签下来，背后是三方都算清楚了账。
第一个现实：算力需求的增长曲线没有放缓迹象。Claude Sonnet 5发布后需求暴增，Mythos和Fable等更大模型的算力需求更是天文数字。CryptoBriefing今年6月的一份分析估算，OpenAI、Anthropic、xAI三家合计消耗了全球AI算力的21%。模型训练的算力需求大约每3-4个月翻一番，推理需求随着用户增长线性甚至超线性上升。Anthropic的CEO Dario Amodei多次公开表示，未来几年AI算力需求会增长几个数量级。20年的合约不是激进，是保守——谁知道2046年的AI需要多少电？
第二个现实：数据中心的建设周期和电力接入周期已经长到令人发指。在美国，一个超大规模数据中心从选址、买地、获得环境许可、接入电网、建成投用，平均需要3-5年。如果是核电或者新型清洁能源，时间更长。你今天发现算力不够了，现建数据中心是来不及的，必须提前5-10年规划。20年的长约给了TeraWulf足够的确定性去融资、买地、接电、建设——银行愿意给一个有20年合同的项目放贷款，电网公司愿意为一个有20年合同的负载扩容线路，地方政府愿意为一个有20年合同的雇主提供税收优惠。
"当你签的不是5年而是20年的合同，你买的不再是服务器时间，你买的是未来20年电力的看涨期权。"—— 一位基础设施投资人的判断
第三个现实：AI公司的竞争维度已经从"谁的模型强"升级到了"谁的基础设施稳"。今年5月，SpaceX把Colossus 1数据中心的全部算力签给了Anthropic，Bloomberg Intelligence估算这笔合作到2028年可能带来超过500亿美元收入；6月，Google承诺5年5GW给Anthropic；7月，TeraWulf的20年合约落地。你会发现一个有趣的现象：Anthropic同时在和Google、Amazon、SpaceX、TeraWulf签约，分散在不同的供应商、不同的地理位置、不同的电力来源。这不是浪费，这是算力层面的"不要把鸡蛋放在一个篮子里"——任何一个供应商出问题（芯片短缺、电力中断、政策变化），都不能让Claude停服。
Meta卖算力：卖铲子的人越来越多
Anthropic疯狂锁算力的同时，另一个标志性事件正在发生：AI公司开始从"算力买家"变成"算力卖家"。
7月1日，Bloomberg报道Meta正在筹建云业务，打算出售过剩的AI算力和模型访问权。消息一出，Meta股价暴涨超过10%。截至第一季度末，Meta已承诺未来几年在AI基础设施上投入1829亿美元——路易斯安那州2250英亩的超大规模园区、俄亥俄州"曼哈顿大小"的吉瓦级数据中心。但现实是：Meta AI和Llama都没有产生显著的外部收入，Muse Spark模型发布时间一再推迟。既然自己用不完，那就租出去。
这不是Meta的原创。今年5月，SpaceX就把Colossus数据中心的算力签给了Anthropic，随后又和Google、Reflection AI签了类似协议。CoreWeave这种"AI算力新云"已经上市，Nebius也在欧洲快速扩张。现在Meta这个级别的巨头入场卖算力，市场格局会发生显著变化。消息公布后，CoreWeave和Nebius股价分别下跌10.8%和12.4%——市场担心Meta这个新玩家会引发价格战。
但从更深层看，Meta卖算力标志着AI算力市场进入了一个新阶段：算力正在从"稀缺资源"变成"可交易的大宗商品"。当自建算力的公司发现自己用不完、当专门做算力租赁的公司上市、当20年的长约成为常态——这说明算力市场正在从"卖家说了算"的短缺时代，转向"有长期合约就有稳定供给"的大宗商品时代。
微软的动作也值得关注。7月2日微软宣布成立Microsoft Frontier Company，投入25亿美元、6000名工程师做AI部署服务。两天后，微软启动了2026年最大规模裁员——近5000人被裁，Xbox部门当天就走了1600人。一边花25亿招6000人做AI，一边裁5000人——这不是矛盾，这是新旧动能转换：微软在把资源从传统业务（游戏主机、传统软件销售、本地咨询）迁移到AI基础设施和AI服务上。
终局判断：AI正在变成重资产行业
把这几条新闻放在一起看，一个清晰的产业拐点正在浮现。
过去三十年，科技行业的主流叙事是"轻资产"——互联网公司不需要拥有工厂、不需要拥有矿山、甚至不需要拥有办公室，几台服务器、一群程序员、一个好idea就能改变世界。Google、Meta、字节跳动，都是轻资产模式的典范。
AI正在把这个叙事彻底砸碎。
训练和运行一个前沿大模型需要什么？需要几十亿上百亿美元买GPU，需要吉瓦级的数据中心，需要几十上百兆瓦的电力供应，需要几年的建设周期，需要和电网公司、地方政府、芯片厂商、云服务商签一揽子长期协议。这已经不是互联网创业的逻辑了，这是19世纪修铁路、20世纪建核电站的逻辑——重资产、长周期、高壁垒、赢家通吃。
Anthropic的20年合约是一个标志性事件。它意味着：
第一，AI创业的门槛会高到难以想象。2023年你还可以靠几千万美元融资训练一个不错的开源模型；2026年，没有几十亿美元你连入场券都拿不到。未来能玩得起前沿大模型的玩家，可能一只手数得过来：微软/OpenAI、Google/Anthropic、Amazon、Meta、字节、腾讯、百度——其他公司要么做垂直应用，要么做开源小模型，要么给这些巨头打工。
第二，算力的金融属性会越来越强。20年190亿美元的合同本身就是一种金融资产——它可以被抵押、被证券化、被交易。未来你可能会看到算力期货、算力ETF、甚至算力的二级市场交易。就像石油期货一样，算力也会成为一种可交易的大宗商品。
第三，AI的地理分布会被电力成本重塑。肯塔基州为什么被选中？因为那里有便宜的电力、有充足的土地、有友好的营商环境。未来AI数据中心会集中在电力便宜（水电、风电、核电丰富）、气候凉爽（散热成本低）、政治稳定的地区。美国的肯塔基、得克萨斯、华盛顿州，北欧的挪威、瑞典，中东的沙特、阿联酋，中国的贵州、内蒙古、宁夏——这些地方会成为AI时代的"油田"。
当然，这里面也有风险。20年是一段很长的时间，谁知道2036年、2046年的计算架构是什么样？如果量子计算或者光子计算在10年内商业化，今天的GPU数据中心可能会变成昂贵的废铁。但Anthropic和TeraWulf都愿意赌——赌的是在可见的未来，基于GPU/TPU的深度学习架构仍然是主流，电力仍然是AI最核心的约束。
19世纪的铁路公司签了99年的土地租约，20世纪的电力公司签了30年的购电协议，21世纪的AI公司签了20年的数据中心合约。每一次技术革命，最终都会落到地上——落到土地上、电线上、水泥建筑里。AI不再是飘在云端的概念，它正在变成一个需要消耗巨量电力、占据大片土地、雇佣数万工程师的实体产业。
当你下次打开Claude聊天的时候，记得你输入的每一个token背后，都有肯塔基州某个小镇上的一台服务器在消耗电力。那台服务器所在的数据中心，是一份20年合同的产物。AI的未来，已经被写进了这些长期合约里。
明天见。
401MW.
What does that number mean? A standard mid-sized nuclear plant clocks in around 1000MW. On July 6, U.S. digital infrastructure firm TeraWulf announced a 20-year data center lease with Anthropic — Anthropic will occupy the Justified Data campus in Hawesville, Kentucky, supporting roughly 401MW of critical IT load, projected to generate $19 billion (~129.2 billion yuan) in full-cycle revenue for TeraWulf. This isn't your average cloud services contract. It's the longest compute lock-in in AI industry history.
Twenty years. Let that sink in. When this contract expires, today's college freshmen will be pushing 40. The iPhone has only existed for 19 years. Google itself is only 28 years old. An AI company founded less than five years ago just locked itself into an infrastructure deal longer than most tech companies have been alive.
The Compute Contract Triple Jump: From Annual Deals to 20-Year Lock-Ins
Look back at the history of AI compute procurement, and you'll see a clear timeline of ever-lengthening contracts.
From 2023 to 2024, LLM companies mostly signed annual cloud contracts. OpenAI's deal with Microsoft was initially renegotiated year by year; everyone was still estimating compute needs at the level of "how many H100s do we need this year." Buying GPUs back then was like grocery shopping — grab 10,000 this year, add more next year if needed.
In 2025, contracts stretched to 3–5 years. OpenAI signed a multi-year supercomputer deal with Microsoft; Google committed 5GW (5000MW) of compute to Anthropic over five years, coming online starting in 2027. Five years is roughly the cycle from planning a data center to running it at full capacity. That's when AI companies started to realize: compute isn't a utility you buy as you go — it's a strategic resource you need to lock in years ahead of time.
In July 2026, Anthropic stretched that timeline to 20 years. This isn't tech company procurement logic anymore — it's power company, railroad company, water utility logic. Name one internet company that signed a 20-year server room lease. You can't. But utilities signing 20-year power purchase agreements for hydro dams? Routine.
The details are telling. The campus is being developed in phases, with initial operations expected in H2 2027 and maximum capacity reached by early 2028. That means Anthropic signs now and waits over a year for the first watt of power. Meanwhile, in the same announcement, TeraWulf said it's selling its entire 50.1% stake in the Abernathy, Texas joint venture (168MW) to partner Fluidstack. This isn't expansion — it's going all-in. Sell off other projects, concentrate every resource on serving one customer: Anthropic.
Why 20 Years? Three Brutal Realities
Anthropic isn't a charity, and neither is TeraWulf. A 20-year, $19 billion deal only happens when all sides have crunched the numbers.
Reality one: The compute demand curve shows zero signs of slowing. Demand exploded after Claude Sonnet 5 launched; even larger models like Mythos and Fable have astronomical compute requirements. A June analysis by CryptoBriefing estimated that OpenAI, Anthropic, and xAI together consume 21% of global AI compute. Training compute needs roughly double every 3–4 months; inference demand grows linearly or even superlinearly as user bases expand. Anthropic CEO Dario Amodei has repeatedly said publicly that AI compute demand will grow by orders of magnitude in the coming years. A 20-year contract isn't aggressive — it's conservative. Who knows how much power AI will need in 2046?
Reality two: Data center construction and grid interconnection timelines have become absurdly long. In the U.S., a hyperscale data center takes an average of 3–5 years from site selection, land purchase, environmental permits, grid hookup, to going live. Nuclear or new clean energy? Even longer. If you realize today you don't have enough compute, building a data center from scratch won't save you — you need to plan 5–10 years out. A 20-year contract gives TeraWulf the certainty to finance, buy land, connect to the grid, and build. Banks will lend to a project with a 20-year contract. Grid operators will expand capacity for a load with a 20-year contract. Local governments will offer tax breaks to an employer with a 20-year contract.
"When you sign a 20-year contract instead of five, you're not buying server time anymore — you're buying a call option on the next 20 years of electricity."— An infrastructure investor's assessment
Reality three: The competitive dimension for AI companies has shifted from "who has the best model" to "who has the most reliable infrastructure." In May, SpaceX signed over all compute from the Colossus 1 data center to Anthropic — Bloomberg Intelligence estimates the deal could generate over $50 billion by 2028. In June, Google committed 5GW over five years to Anthropic. In July, TeraWulf's 20-year deal landed. Notice a pattern? Anthropic is signing with Google, Amazon, SpaceX, and TeraWulf simultaneously — spread across different vendors, geographies, and power sources. This isn't waste. It's compute-level "don't put all your eggs in one basket." If any one supplier hits trouble — chip shortages, power outages, policy shifts — Claude must not go down.
Meta Sells Compute: The Shovel Sellers Multiply
While Anthropic is hoovering up compute, another landmark shift is underway: AI companies are going from compute buyers to compute sellers.
On July 1, Bloomberg reported that Meta is building a cloud business to sell excess AI compute and model access. Meta's stock jumped over 10% on the news. As of Q1, Meta had committed $182.9 billion to AI infrastructure over the coming years — a 2,250-acre hyperscale campus in Louisiana, a "Manhattan-sized" gigawatt-scale data center in Ohio. The reality? Neither Meta AI nor Llama has generated meaningful external revenue; the Muse Spark model has been delayed repeatedly. If you can't use it all yourself, rent it out.
Meta didn't invent this move. In May, SpaceX signed Colossus data center compute to Anthropic, then struck similar deals with Google and Reflection AI. "New AI cloud" players like CoreWeave have gone public; Nebius is expanding rapidly in Europe. Now a giant of Meta's caliber enters the compute resale market, and the landscape shifts dramatically. After the news broke, CoreWeave and Nebius shares dropped 10.8% and 12.4% respectively — the market fears a price war triggered by this new entrant.
But look deeper: Meta selling compute marks a new phase for the AI compute market — compute is shifting from a scarce resource to a tradable commodity. When companies that build their own compute find they can't use it all, when pure-play compute rental firms go public, when 20-year contracts become the norm — the compute market is moving from a seller's-market shortage era to a commodity era where long-term contracts guarantee stable supply.
Microsoft's moves deserve attention too. On July 2, Microsoft announced the Microsoft Frontier Company, pouring $2.5 billion and 6,000 engineers into AI deployment services. Two days later, Microsoft launched its biggest layoff round of 2026 — nearly 5,000 people cut, with 1,600 leaving the Xbox division in a single day. Spend $2.5 billion hiring 6,000 for AI while cutting 5,000 jobs — this isn't a contradiction; it's a pivot from legacy to new growth engines. Microsoft is shifting resources from traditional businesses (consoles, legacy software sales, on-prem consulting) to AI infrastructure and AI services.
Endgame: AI Is Becoming a Heavy-Asset Industry
Put these stories together, and a clear inflection point comes into focus.
For the past three decades, the dominant tech industry narrative was "asset-light" — internet companies didn't need factories, mines, or even offices; a few servers, a bunch of programmers, and a good idea could change the world. Google, Meta, ByteDance — all poster children for the asset-light model.
AI is smashing that narrative to pieces.
What does it take to train and run a frontier LLM? Tens of billions of dollars in GPUs. Gigawatt-scale data centers. Hundreds of megawatts of power. Years of construction. A web of long-term deals with grid operators, local governments, chipmakers, and cloud providers. This isn't internet startup logic anymore. It's 19th-century railroad logic, 20th-century nuclear plant logic — heavy assets, long cycles, high barriers, winner takes all.
Anthropic's 20-year contract is a landmark moment. It means three things:
First, the barrier to entry for AI startups is going to become unimaginably high. In 2023 you could train a decent open-source model on tens of millions in funding. In 2026, without billions, you can't even buy a ticket to the game. The number of players who can afford frontier models might fit on one hand: Microsoft/OpenAI, Google/Anthropic, Amazon, Meta, ByteDance, Tencent, Baidu. Everyone else either builds vertical apps, works on small open-source models, or works for these giants.
Second, compute will become increasingly financialized. A 20-year, $19 billion contract is itself a financial asset — it can be collateralized, securitized, traded. In the future you might see compute futures, compute ETFs, even a secondary market for compute. Just like oil futures, compute will become a tradable commodity.
Third, the geography of AI will be reshaped by power costs. Why Kentucky? Cheap electricity, abundant land, business-friendly policies. Future AI data centers will cluster where power is cheap (rich in hydro, wind, nuclear), climates are cool (lower cooling costs), and politics are stable. Kentucky, Texas, Washington State in the U.S.; Norway and Sweden in the Nordics; Saudi Arabia and the UAE in the Middle East; Guizhou, Inner Mongolia, and Ningxia in China — these places will become the "oil fields" of the AI era.
Of course, there are risks. Twenty years is a long time. Who knows what computing architecture will look like in 2036 or 2046? If quantum or photonic computing commercializes within a decade, today's GPU data centers could become expensive scrap metal. But both Anthropic and TeraWulf are willing to make the bet — that for the foreseeable future, GPU/TPU-based deep learning architectures remain dominant, and power remains AI's most fundamental constraint.
19th-century railroads signed 99-year land leases. 20th-century utilities signed 30-year power purchase agreements. 21st-century AI companies sign 20-year data center contracts. Every technological revolution eventually hits the ground — lands on soil, wires, concrete buildings. AI is no longer a concept floating in the cloud; it's becoming a physical industry that consumes massive amounts of power, occupies vast tracts of land, and employs tens of thousands of engineers.
Next time you open Claude to chat, remember: every token you type is being processed by a server in some small Kentucky town, consuming electricity. The data center housing that server is the product of a 20-year contract. AI's future has already been written into these long-term deals.
See you tomorrow.
Anthropic · TeraWulf · 190亿美元 · 20年算力合约 · 401MW · AI军备竞赛 · 算力大宗商品 · Meta卖算力 · 重资产AI · 电力基础设施
Anthropic · TeraWulf · $19 billion · 20-year compute contract · 401MW · AI arms race · compute as commodity · Meta sells compute · heavy-asset AI · power infrastructure
