The AI Race Moved to the Floor

CoreWeave had a $100 billion revenue backlog and the largest hyperscalers as clients ^[1]. Then its biggest customer decided to compete with it. In early July, Meta — which accounted for $21 billion of CoreWeave's contracted revenue — launched Meta Compute, a service that sells excess GPU capacity directly to the market ^[2].

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

The move turned CoreWeave's largest client into its newest competitor, and the market delivered its verdict with speed and precision: over two weeks this July, it erased half of CoreWeave's market value, a $49 billion deletion ^[3][4]. The crash was not a story about one company's execution. It was the market pricing in a structural fact the AI industry has been hurtling toward for months. CoreWeave is a neocloud — an intermediary that rents data-center space and resells GPU access to companies that need AI compute. It carries $50.8 billion in total liabilities against $1.2 billion in net losses, and it relies heavily on renting data center space rather than owning it ^[3]. When the client that supplied a fifth of its business built its own infrastructure and started selling the surplus, the market recognized that an intermediary without a physical layer has no durable moat. Ownership of the physical layer has become the defensible position. The same month that CoreWeave cratered, the model layer above it continued its own race to the bottom. In May, DeepSeek permanently cut prices on its V4-Pro model by 75%, bringing its cost to roughly one-twelfth of what OpenAI charges for GPT-5.5 on equivalent tasks ^[5]. The price cut was enabled by something specific: DeepSeek runs on Huawei's Ascend 950 processors, domestic Chinese silicon that insulates it from export controls and from the margin structure of US chip supply ^[5].

It is an efficiency gain being passed through. — Sanchit Vir Gogia

Then in July, Zhipu AI released GLM-5.2, an open-weights model that matches frontier performance at a fraction of the cost ^[6].

We now have a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic. — David Sacks

The model layer is not standing still. GPT-5.6 launched in July, Claude Opus 4.8 in May, and the frontier continues to advance ^[7][8]. But capability is no longer scarce. When a Chinese lab can match US frontier performance at a fraction of the price, and when open-weights models erase the distribution advantage of proprietary systems, the model itself stops being a moat. It becomes a commodity — widely available, rapidly replicable, and priced accordingly. The scarcity has migrated downstairs. The physical layer that runs AI — data centers, power generation, cooling systems, semiconductor fabrication — is now the binding constraint on everything above it. The global data center market faces a 12-gigawatt capacity deficit because supply chains for engineering labor, cooling equipment, and electrical components cannot keep pace with demand ^[9].

Demand for data centers continues to outpace supply, with hyperscaler capex accelerating and chip volume forecasts implying GWs of capacity ahead of feasible data center delivery. — Jefferies Group

The numbers that define this layer are not software numbers. Nvidia projects that global data center capital expenditure will reach $3 to $4 trillion annually by 2030, with the four largest hyperscalers spending over $650 billion in 2026 alone ^[10]. Amazon is investing $200 billion in data center capex this year; AWS now generates 59% of the company's total operating profits, making physical infrastructure the profit center of one of the world's largest corporations ^[11]. US data centers are projected to consume between 9.5% and 15% of total national electricity by 2030, with Chase Bank raising its capacity growth forecast to 138 gigawatts ^[12].

the scale and growth of computational demand more than offset these efficiency gains, leading to continued increases in absolute electricity consumption. — Lawrence Berkeley National Laboratory

The supply chain is straining at every joint. Nine US trade groups warned that the AI data center buildout is causing a critical memory chip shortage that threatens production of consumer electronics, vehicles, and medical devices. SK hynix's chairman estimates the shortage could persist until 2030, because semiconductor fabrication plants take years to build ^[13].

an urgent imbalance in the market for memory chips could lead to significant and sustained near-term price increases for American households and disrupt critical U.S. supply chains. — Alliance for Automotive Innovation

This is not a temporary bottleneck. It is the shape of the new competitive landscape. The actors who recognized it earliest are not the model labs — they are the hyperscalers, the nations, and the chipmakers who are building the physical environment while the software layer commoditizes above them. Japan moved first and most explicitly. In mid-July, it announced the world's first national "physical AI" infrastructure with Nvidia — a 140-megawatt AI factory running 27,500 Rubin GPUs, backed by ¥1 trillion through 2030 ^[14]. Jensen Huang framed it in terms that left no ambiguity about where he believes the value has migrated.

The next frontier of AI is in the physical world, and this is a once-in-a-generation opportunity for Japan. — Jensen Huang

China's response is larger in scale and more explicit in its strategic intent. Beijing committed $295 billion over five years to a nationwide AI data center network, with a mandate that 80% of core technology — including AI chips — be domestically sourced ^[15]. The policy is designed to escape dependence on US hardware, and it is already producing results: DeepSeek's price advantage flows directly from Huawei's Ascend chips, not from algorithmic breakthroughs alone. Moonshot AI's president made the dynamic explicit when describing the constraints that shaped his lab's approach.

We knew we didn’t have the luxury to simply scale up compute. — Yutong Zhang

India's sovereign AI push was triggered by a more immediate shock. When a US export-control directive forced Anthropic to abruptly cut off foreign access to its models, the Indian government and its AI ecosystem recognized that access to someone else's infrastructure is not the same as owning it. Sarvam AI, one of India's leading labs, issued a warning that has become the unofficial motto of the sovereign AI movement ^[16].

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

India has since moved on multiple fronts: it inaugurated a Jabil manufacturing plant in Maharashtra to domestically produce AI data center components, allocated 88,000 GPUs to startups, and funded 20 foundational models through the IndiaAI Mission ^[17]. The pattern is extending from operating data centers to manufacturing the physical components that go inside them. Europe is following the same logic. Foxconn, Bull, and Schneider Electric are building AI factory supply chains across France and the Czech Republic using Nvidia's Vera Rubin platform ^[18].

Foxconn is proud to partner with Bull and NVIDIA to help build the foundation for AI factories, sovereign AI and next-generation data center infrastructure in Europe. — Hon Hai Precision Industry

Alibaba Cloud launched a Paris region marketed as "sovereign, secure, and intelligent solutions" — a Chinese cloud provider building physical infrastructure inside Europe to meet European data sovereignty requirements ^[19]. Even SoftBank, a Japanese firm, is funding a $1.5 trillion AI data center campus in Ohio on a former uranium enrichment site, with 18 small modular nuclear reactors and a $33 billion gas plant ^[20].

We are embarking on the next era of American dominance. — Tim Walsh

The AI buildout has become, literally, a nuclear power plant business — AtkinsRéalis and Nvidia are now partnering to design nuclear-powered AI factories using digital twins before physical construction begins ^[21]. None of this means the model layer has stopped mattering. Frontier models continue to ship monthly, and the software race is real. Edge inference — running AI on local devices rather than centralized cloud — may decentralize some workloads, and open-weights models like GLM-5.2 could shift compute from massive data centers to smaller, distributed hardware ^[22]. Community opposition is also a genuine constraint: 75 AI data center projects worth $130 billion were blocked or delayed in the first quarter of 2026, with opposition groups doubling to 833 across 49 states and multiple cities passing permanent bans or moratoriums ^[23][24]. But these forces constrain the pace of the buildout, not its direction. The capital flows tell the story. The market looked at CoreWeave — a company with a $100 billion backlog, soaring demand, and no physical assets it controlled — and cut it in half the moment its largest client built its own infrastructure. The same market is directing trillions toward the companies and countries that own the concrete, the silicon, and the power plants. The AI race did not stop. It moved to the floor, and the actors who saw it — hyperscalers, nations, DeepSeek's hardware-enabled pricing — are the ones building while the intermediary in the middle gets cut in half.