AI到底有多烧钱？过去的答案是「多少万块GPU」，现在的答案变成了「多少吉瓦」。
8月17日，OpenAI正式宣布加入PORTS-Pike项目——在美国俄亥俄州Pike县建设世界级AI基础设施园区。整个项目锁定约8GW IT容量，由SB Energy建设并运营，OpenAI签了20年租约，英伟达提供信用担保并成为独家算力供应商。
8GW是什么概念？大约是一座超大型工业基地的用电负荷。首批800MW预计2028年投用，后续还需要新建电厂、输电线路和配套设施。
150万块GPU的含义
按照英伟达的测算，初始4.25GW部署，每一代AI工厂系统对应约150万块GPU，可能带来1500亿到2000亿美元的收入机会。如果后续扩展到全部8GW甚至16GW，对应的长期计算机会约6000亿美元。
注意，这不是说OpenAI今天已经签了6000亿的订单。这是英伟达按20年规划周期、多代GPU升级测算出的长期市场机会。第一代装进去，几年后升级第二代，再升级第三代——AI数据中心不是建一次用20年，而是每隔几年就换一次「心脏」。
很多人对10万、20万卡的集群已经觉得规模巨大，现在行业开始讨论百万级GPU基础设施。这是量级的跃迁。
英伟达角色的质变
这个项目里最值得关注的，不是又卖了多少GPU，而是英伟达角色的根本变化。
过去它只是GPU供应商，一手交钱一手交货。现在它越来越深地介入：芯片供应、网络、软件、数据中心设计、资本投资，甚至项目融资担保。英伟达向SB Energy投资了15亿美元，同时为前4.25GW提供最高1050亿美元的租赁残值担保。
为什么？因为如果未来AI算力真的是一个数千亿美元的市场，对于英伟达而言，最重要的事情已经不是「让客户买我的GPU」，而是确保客户真的有钱、有电、有数据中心把GPU买回去。
Jensen Huang造了一个新词：LPS——Land（土地）、Power（电力）、Shell（建筑外壳）。这三样东西，已经取代了芯片本身，成为AI扩张的真正瓶颈。
明天见。
How expensive is AI to build? Not long ago, the answer was 'how many tens of thousands of GPUs.' Today the unit is gigawatts.
On August 17, OpenAI formally joined the PORTS-Pike project — a world-class AI infrastructure campus in Pike County, Ohio. The entire development locks in approximately 8 gigawatts of IT capacity. SB Energy builds and operates it, OpenAI signed a 20-year lease, and NVIDIA provides credit support as the exclusive compute provider.
8GW is roughly the electrical load of a massive industrial complex. The first 800MW comes online in 2028, with new power plants, transmission lines, and supporting infrastructure still to come.
What 1.5 Million GPUs Actually Means
By NVIDIA's estimates, the initial 4.25GW deployment translates to roughly 1.5 million GPUs per system generation, representing $150–200 billion in revenue opportunity. If the project expands to the full 8GW or even 16GW down the line, the long-term compute opportunity reaches approximately $600 billion.
To be clear: OpenAI hasn't signed a $600B order today. This is NVIDIA's long-term math across a 20-year planning horizon with multiple GPU generations. You install generation one, then upgrade to generation two a few years later, then generation three. AI data centers aren't built once and used for 20 years — they get a new 'heart' every few years.
People already find 100K–200K GPU clusters enormous. Now the industry is talking about million-GPU infrastructure. That's a different order of magnitude.
NVIDIA's Role Has Fundamentally Changed
The most interesting part of this project isn't how many more GPUs got sold. It's how NVIDIA's role has transformed.
It used to be just a GPU supplier — cash on the barrelhead, chips in a box. Now it's getting deeper into everything: silicon supply, networking, software, data center design, capital investment, even project financing guarantees. NVIDIA invested $1.5 billion in SB Energy, and it's backing the first 4.25GW with up to $105 billion in residual value guarantees.
Why? Because if AI compute is genuinely a multi-hundred-billion-dollar market, NVIDIA's biggest priority isn't 'getting customers to buy GPUs' — it's making sure customers can actually afford to build, power, and house those GPUs.
Jensen Huang coined a new acronym: LPS — Land, Power, Shell. Those three things, not chips, have become the real bottleneck of AI expansion.
See you tomorrow.
