Compute Scarcity Crisis

Jane Street has agreed to spend roughly $6 billion on CoreWeave’s GPU cloud and to take about $1 billion more in equity—an arrangement that buys capacity, not just compute cycles Jane Street signs $6 billion AI cloud deal with Coreweave, boosts stake Jane Street Invests $1 Billion in CoreWeave, Boosts Spending Plans. That deal, one of several large commitments this week including an expanded Meta arrangement reportedly worth about $21 billion and a claimed CoreWeave backlog of $88 billion, makes plain what was becoming obvious in the background: access is the scarce product in high-end AI CoreWeave Has a Massive $88 Billion Revenue Backlog. Here's Why the Stock Could 10x in 5 Years. CoreWeave Just Signed a $21 Billion Deal With Meta. Here's Why This Stock Might Finally Be Turning the Corner.. The commercial logic is simple and specific. Frontier models are easy to copy; guaranteed inference capacity is not. Sophisticated buyers are buying time on premium GPUs, datacentre racks, and the long-term contracts that ensure inference runs when needed. In market terms, customers are paying a premium for certainty rather than haggling for marginal discounts. That shifts the prize. The lab with the neatest demo no longer has first claim on market power; the vendor that can lock in chips, power and permits does. The hardware chain is now entangled with geopolitics. ASML beat Q1 expectations and raised its 2026 sales guidance this week, yet its shares fell about 5% as investors fretted over tighter China export rules—echoing the stalling Nvidia and AMD AI-chip approvals to China from bureaucratic bottlenecks like BIS staffing shortages—an object lesson that lithography tools are political as well as industrial ASML stock sinks amid tightening China restrictions despite strong earnings, guidance. Upstream policy can throttle downstream capacity. Local politics is constraining buildout too, compounding the water access strains from yesterday's record-low snowpack across western states. Public opposition to datacentres—over electricity, water, land use and transmission lines—is rising in the United States and is already affecting where and how fast capacity comes online The public sours on AI and data centers as Anthropic, OpenAI look to IPO and tech keeps spending. Even OpenAI, which has publicly touted aggressive capacity plans, stepped back from an offtake for roughly half of a 230 MW Narvik site; Microsoft took over the commitment instead, underscoring how hard it is to make bankable supply pledges for huge facilities OpenAI pulls back from Stargate Norway data center deal as Microsoft takes over. The commercial ripples are visible in product strategy and developer communities. Anthropic’s reported move from flat subscriptions toward usage-based enterprise pricing—some heavy users reportedly seeing bills double or triple—looks less like experimentation and more like passing scarce compute costs to customers Demand Surges, Computing Power Strained! Anthropic Adjusts Claude Enterprise Pricing to Usage-Based Model. At the same time, open-source and systems projects are accelerating optimisations: UC Berkeley’s vLLM has drawn large community interest (tens of thousands of GitHub stars) for memory-efficient serving, building on the scalable OCR demos that showed how it slashes inference costs, and experimental techniques such as KVSwap promise 1.8x–4.1x throughput gains by tiering KV caches to local storage—evidence that engineers are treating inference thrift as a product requirement, not a savings exercise.[[1]](https://x.com/i/status/2044426036763599280) (AI Post Transformers; internal notes on KVSwap). Risk is now a procurement argument as well as an engineering one. Reports that powerful models have been withheld or tightly scoped—access to some frontier systems limited to a few dozen organisations under controlled programmes—signal that vendors and regulators view deployment as a systemic-risk decision, not simply a marketing event (podcast reporting; Project Glasswing access limited to about 40 organisations). The implication for buyers and builders is stark: resilience now requires multi-provider portability, mixed-tier serving, aggressive quantisation and routing, and a product design that treats cost-per-useful-output as a feature. The decisive contest in AI is therefore less about model cards than about industrial logistics: who can secure chips, datacentres, power contracts and social licence to run them. The winners will be the firms that marry model engineering to procurement, power-market strategy and local politics—and accept that, for now, access is the most valuable commodity.