Jane Street은 AI 컴퓨팅 투자에 190억 달러를 배정하는 한편, Crusoe는 30억 달러의 자금 조달 라운드를 성공적으로 마감했으며, 금융 자본이 전통적인 연구소 지출을 점차 앞지르고 있다.
When Bloomberg reported on September 3 that Denver-based AI infrastructure company Crusoe had closed a $3 billion Series F at a $30 billion post-money valuation, the headline figure was striking but not entirely new — TechTimes had previously reported those talks in July. What distinguished this deal, and what most coverage buried past the funding announcement, was the buyer attached to Crusoe’s largest customer contract: Jane Street Group. The quantitative trading firm committed approximately $13 billion in cloud spending to Crusoe over five years. That same week, Jane Street also led a $1.5 billion equity round in rival AI infrastructure startup FluidStack. Combined with Jane Street’s $6 billion cloud commitment and $1 billion equity stake in CoreWeave secured in April, a firm that has never sold an AI product now holds more contracted AI compute than most frontier model developers will ever consume. It has emerged as the AI infrastructure sector’s most consequential non-hyperscaler backer.
Jane Street Rewrites the AI Infrastructure Demand Picture
For much of the AI infrastructure boom, demand was driven by model developers and hyperscalers—OpenAI, Meta, Anthropic, Google, and Microsoft—competing for scarce GPU capacity to train progressively larger models. Jane Street’s spending pattern in 2026 fundamentally alters that narrative. The trading firm’s AI infrastructure commitments now total roughly $21.5 billion: $6 billion in cloud capacity from CoreWeave, a $1 billion equity position in CoreWeave at $109 per share, approximately $13 billion in Crusoe cloud capacity over five years, and the $1.5 billion FluidStack equity investment it led this week. Confirmed across TechCrunch’s Series F reporting, CoreWeave’s official press release, and Crunchbase’s weekly funding roundup, this total exceeds the GPU infrastructure commitments of most AI laboratories. Jane Street requires compute for reasons structurally distinct from those of a model developer.
Quantitative trading is among the most computationally intensive enterprises globally. Firms like Jane Street deploy machine learning to generate trading signals, simulate market scenarios, optimize execution algorithms, and stress-test portfolio models in real time. This simultaneous, large-scale processing compounds their compute appetite with each new generation of models. The critical distinction from a standard AI lab lies in competitive cadence: while a frontier model laboratory trains a new model every few months, a quantitative trading firm refines and re-trains models daily or hourly. Each marginal improvement in prediction accuracy translates directly into captured profit. Faster iteration demands more compute, and dedicated GPU clusters ensure that iteration is never throttled by shared cloud queues or spot-market pricing.
CoreWeave’s official press release announcing the April deal explicitly acknowledged this dynamic, noting that it would provide Jane Street with access to next-generation NVIDIA Vera Rubin chips—at the time among the first commercially available—alongside dedicated connectivity and custom storage configurations. In quantitative trading, securing hardware ahead of competitors is not merely a cost-optimization decision; it is a structural competitive advantage.
Crusoe's Series F: From Talks to Closed, With Named Investors
The closed round aligns structurally with Bloomberg’s earlier reporting: $3 billion raised at approximately a $30 billion post-money valuation, framed as a “pre-IPO” transaction. The newly confirmed leadership includes Atreides Management and Valor Equity Partners co-leading the financing, with participation from Mubadala Capital—the alternative asset management arm of Abu Dhabi’s $302 billion sovereign wealth fund, Mubadala Investment Company—as noted in TechCrunch’s Series F reporting. The “pre-IPO” designation carries its own signal, typically implying a public listing within 12 to 24 months.
Atreides Management is a Boston-based investment firm managed by Gavin Baker, known for concentrated bets on technology companies at the growth stage. It has co-invested with Valor Equity Partners through the dedicated Valor Atreides AI Fund, a vehicle structured specifically for AI infrastructure and computing companies. Valor Equity Partners, led by Antonio Gracias, initially backed Crusoe in its Series E round in October 2025. Its decision to reinvest at a tripled valuation signals strong conviction in the company’s trajectory rather than a forced-hand continuation.
Mubadala Capital’s participation ties the round to a broader global pattern of sovereign capital accelerating into AI infrastructure. Abu Dhabi’s sovereign wealth ecosystem—which collectively manages approximately $1.7 trillion across ADIA, Mubadala, and ADQ—has positioned AI infrastructure as a core strategic allocation rather than a purely financial bet. Mubadala had also participated in Crusoe’s Series E, marking its second consecutive commitment to the company.
The $30 billion post-money valuation implies a pre-money figure of roughly $27 billion, representing a nearly threefold increase from the $10 billion valuation established in October 2025, according to TechCrunch’s Series F reporting. Total equity raised by Crusoe now exceeds $7 billion, per Crunchbase, positioning it as the most heavily capitalized private AI infrastructure company not yet publicly traded.
From Flared Gas to AI Factories: How Crusoe Built the Plumbing Finance Now Depends On
Crusoe’s founding story—capturing natural gas flared and wasted at remote oil wells, converting it into cheap electricity, and running compute on that power—is well known to anyone following the neocloud sector. What is less widely understood is why the company’s current architecture positions it structurally better than most neocloud rivals to secure large, single-tenant, dedicated-compute contracts like Jane Street’s $13 billion agreement.
Standard GPU cloud operators—the category often termed neoclouds—acquire or lease GPU hardware, rack it at colocation facilities, and sell compute time at rates 40 to 85 percent below hyperscaler pricing for equivalent silicon. While highly appealing to buyers, this model yields narrow economics for operators. Prior TechTimes analysis cited McKinsey’s estimate that bare-metal-as-a-service businesses post gross profit margins of only 14 to 16 percent after accounting for labor, power, and depreciation—lower than many non-technology retailers. Furthermore, each new chip generation typically halves the rental price of older hardware within five years, and revenue often concentrates in one or two customers.
Crusoe’s model is structurally distinct. The company owns or co-develops the land, power generation assets, and data center buildings at each campus, then layers GPU hardware and cloud software on top. It manufactures prefabricated power and data center modules at its own facilities in Colorado, Oklahoma, and Louisiana—a capability it calls its Spark Factory—and ships them ready for field installation, as detailed in Crusoe’s June 2026 contracted capacity announcement. Rather than waiting for sequential power-and-construction negotiations, the company develops both tracks simultaneously, compressing the timeline from site selection to energized capacity.
At its flagship 1.2-gigawatt campus on the Lancium Clean Campus in Abilene, Texas—built for Oracle and serving OpenAI’s Stargate project—the physical site moved from groundbreaking to energized buildings in under one year, as documented in prior TechTimes coverage. The power architecture illustrates the underlying cost logic: the Texas Commission on Environmental Quality permitted Crusoe to operate ten simple-cycle turbines providing approximately 360 megawatts of behind-the-meter generation. Because this electricity is produced and consumed on-site, authorized exclusively for the data center, it bypasses the multi-year interconnection queue that confronts grid-connected competitors. Given that energy represents roughly 60 percent of AI data center operating expenses, structural access to cheaper, dedicated power is not merely a marginal pricing advantage; it is a foundational business model that can be sustained where others cannot.
The campus design itself reflects the extreme density requirements of modern AI hardware. A single rack of NVIDIA H100 SXM5 GPUs draws approximately 56 kilowatts at full load, while newer GB200 NVL72 racks push well beyond that threshold. By comparison, legacy data center racks drew only two to four kilowatts. This thirty-five- to seventy-fold density increase demands a completely different approach to power distribution, cooling, and building layout. CEO Chase Lochmiller has described the facility’s design as a printed circuit board at campus scale: GPU compute wings radiating outward from a central core housing network and storage infrastructure.